Avian Physics

[surreal junco]

So I think that's a bug with 2D picking in general, or an unexpected behavior if it isn't a bug (I could imagine that the point query has a tolerance, but I can't find evidence of that)

From what I'm reading, it looks like a spatial query (specifically SpatialQuery::point_intersections_callback) is documented to find all entities with a collider that contains the given point, but that delegates to the BVH, which as far as I can tell only checks the AABBs in the tree and never checks intersection with the actual collider. I could be wrong on this; I'm still very new to the codebase, but I don't see a direct point-against-collider check anywhere in the call tree starting from the spatial query.

@vestal minnow These are my findings. Let me know if there's any other way I can help, including making a PR (and what the PR should do; is the documentation wrong, or is the code wrong? or are neither wrong and i'm misreading something 🙃)

If you want to try what I have out, I have a branch on my fork with the example shown in the video: https://github.com/rdrpenguin04/avian/blob/rdr-2d-in-perspective-picking/crates/avian2d/examples/picking_2d_in_perspective.rs as well as my other change to the picking plugin

[vestal minnow]

Ah, yeah that's definitely a regression. Pushed a fix here
https://github.com/avianphysics/avian/pull/961

[vestal minnow]

Seems like it works when using both my fix and your perspective fix!

I would appreciate a PR with the perspective fix in src/picking/mod.rs

We can add the example too if you want, no strong opinion there

[surreal junco]

PR has been submitted 😄 https://github.com/avianphysics/avian/pull/962

[vestal minnow]

Thanks! I'll review and merge later today, and make a 0.6.1 patch release with this (and some other fixes) tomorrow

[surreal junco]

Awesome! 😄

[lucid cloud]

Is this issue on anyone's radar? 🥺 https://github.com/avianphysics/avian/issues/946

[vestal minnow]

Ohh sorry I think I missed that one or forgot about it! It looks like the same problem as the very recent issue:
https://github.com/avianphysics/avian/issues/957

If it is that, it should be fixed on the main branch already

And will be fixed in the upcoming patch release

[lucid cloud]

It actually is fixed 🙌 thank you!

Closed the issue ✅

[nimble fiber]

damn you went down a rabbit hole glad y'all were able to figure it out haha

[torn folio]

i really like how this turned out due to me using avian. Raising and lowering the gates using distance joints makes the movement look really really good

the little shake when it reaches the top and botton positions are just wonderful

(this is a model of a furnace I have worked with in real life, and the gates there do the exact same shakes, which is really nice to be able to model with just three ropes and a system)

[nimble fiber]

don't mean to get too off topic for the thread, but your job looks kinda terrifying lol

how tall are those doors in real life?

[torn folio]

like 2m tall maybe? A bit taller than me at least

the scale is very off in the model right now

also this is a furnace, which is why I'm using a scary red glow lol

based on a 40mw submerged arc furnace

[torn folio]

a physics debug view if anyone's curious (distance joints on the gates, prismatic joints on the rods)

[surreal junco]

Alright, my use case is getting more and more... funny.

I'm switching to basically precisely imitating SSBM's Environment Collision Box model. This means that I have four points that move with the character, changing position depending on the animation. These four points in SSBM always formed a kite, but they can generally be any quadrilateral (I don't think being a kite was necessary for any particular logic). I presently use MoveAndSlide, which takes a Collider as input and sweeps it along its velocity vector, making any adjustments to velocity necessary along the way. ... But now I have four colliders.

This gives me three questions:

• What's the most idiomatic way to make a point collider? I could do circle(0.0), rectangle(0.0, 0.0), segment(vec2(0.0, 0.0), vec2(0.0, 0.0)), or a number of other things; which one is subjectively cleanest and/or objectively fastest?
• Can you make a compound collider where you can independently transform the individual parts? I could believe the answer to be no just because this would require a lot of AABB calculation, and it also just wouldn't work well with MoveAndSlide anyway; teleportation in physics engines is bad. Which leads me to...
• What's a good way to make the MoveAndSlide algorithm work with a set of several colliders? A more sane usecase than mine would be move-and-slide-ing a character with a swishing tail or bulky arms or something like that which should push against the geometry.

[nimble fiber]

Lmk if I'm not helping, but why not just use a convex polyline with 4 points? You end up with just one collider so that resolves point #3 and #1, and you aren't dealing with "infinitely thin" colliders and are therefore less likely to have weird issues with characters getting stuck on other infinitely thin surfaces like concave polygons. I know you'd be altering the collider with every state change, to which I say again, AABB/BVH recalcs aren't THAT expensive, they're O(n) or better. So unless you have thousands of characters I wouldn't worry about it and follow the old idiom of profile first optimize second.

Point #2 is still unresolved- Having dealt with a similar but unrelated issue (implementing an HL2 style gravity gun where holding an object adds a physics exception between the holder and the object, which means you can "teleport" it into yourself by picking it up moving it into you and dropping it), you may need to use SpatialQuery to depenetrate your character when their collider changes manually. Though maybe test with MoveAndSlide first, it has two depen steps, I just don't know if they handle colliders that are already penetrating when the step begins.

[surreal junco]

The biggest problem I think is that I need to edit the collider in realtime, which... there doesn't seem to exist a way to do so without creating a fresh collider for each change

Like, even using SharedShape directly only gives me an Arc, and it looks like the code expects no interior mutability, so the best I can do to change the collider is to just make an entirely new collider, which feels significantly worse

(plus, the MoveAndSlide algorithm then has no clue what happened, and then you're right, I'd need to depenetrate manually, which sounds not-fun to do; at that point I may as well make my own move and slide impl because the collider will sometimes be changine frame to frame)

[nimble fiber]

Making your own move and slide isn't that hard, also have done that before. The nice thing is you can just look at the source code of the existing implementation and tweak the parts you need to

[nimble fiber]

I mean, I know it seems really inefficient but have you tried it? I feel like I keep smashing the "premature optimization is the root of all evil" button but I would be shocked if you came anywhere close to dropping a frame. I've generated like, 64^3 colliders in 400ms in Godot before with GDScript without dropping a frame. I bet Avian will handle it lol

[surreal junco]

That's fair; it also just doesn't seem like the right solution intuitively, which is why I'm trying to find another answer

[nimble fiber]

I can see that. Well, I'll shut my yapper and leave room for someone else to make suggestions lol

[surreal junco]

You're good 🙃

[cedar hollow]

huh????? lol

must have a wrong version of glam from something else

[vestal minnow]

try cargo update

[cedar hollow]

yeah 👍 thats what worked

[vestal minnow]

What's the most idiomatic way to make a point collider?
Probably a tiny circle/sphere

Can you make a compound collider where you can independently transform the individual parts?
Yeah, Collider::compound takes a vec of shapes and their positions and rotations, or alternatively you can just add the colliders as child entities and give them different Transforms

What's a good way to make the MoveAndSlide algorithm work with a set of several colliders?
This would require using Collider::compound since it needs one shape (even if it consists of multiple sub-shapes) to use for the shape casts and contact queries

In general though, changing colliders dynamically is kinda wonky and probably won't give very good results since it can make you suddenly clip into geometry with the new shape etc.

You can use a different hitbox and hurtbox every frame and during animation, but changing the actual collider that collides with the environment feels a bit sketchier

But it should probably be fine if you really need it

[surreal junco]

Alright; I'll let you know what happens!

[glossy holly]

I think there was an undocumented change w collider scaling last update, a collider that scaled with transform before no longer does so

[heady oasis]

soo i want to cast a ray with the spacialquery cast_ray and i want to exclude the entity that "is casting the ray"
do I query for the entity or is there some way to fetch this entity from its attstched components?

[ancient quarry]

How is your system determining where to cast the ray from? If you have a query reading something like &GlobalTransform for it, you can add Entity to the QueryData to

[heady oasis]

https://codeberg.org/gobbe/brumm/src/branch/main/src/bil.rs exact code here (raycast at line 136)

either way it querys for the wheel entity and then sends a ray from entities global position

downwards

[little maple]

has anyone else run into this panic?

thread 'main' (504637) panicked at avian/crates/avian3d/../../src/data_structures/stable_vec.rs:334:13:
no element at index 5 in StableVec::index_mut

It started for me on 17d868e1 BVH Broad Phase and Collider Trees (#927)

[vestal minnow]

the release I just published fixes one known BVH crash, but that one might be different

[little maple]

ah, let me try that

[vestal minnow]

most likely different, the fixed one is a None unwrap after spawning/despawning entities in a collision event observer

though I guess there could be other ways it manifests too

[little maple]

it didn't fix it.. but I think it's related to despawning an entity

Encountered a panic in system avian3d::collider_tree::update::update_moved_collider_aabbs<avian3d::collision::collider::parry::Collider>!
Encountered a panic in system avian3d::schedule::run_physics_schedule!
Encountered a panic in system bevy_app::main_schedule::FixedMain::run_fixed_main!
Encountered a panic in system bevy_time::fixed::run_fixed_main_schedule!
Encountered a panic in system bevy_app::main_schedule::Main::run_main!

[little maple]

I rewrote this
https://github.com/avianphysics/avian/blob/v0.6.1/src/collider_tree/update.rs#L996

to be this so it just checks instead

            let trailing_zeros = bits.trailing_zeros();
            let proxy_id = ProxyId::new(i as u32 * 64 + trailing_zeros);

            if let Some(proxy) = tree.proxies.get_mut(proxy_id.index()) {
                let entity = proxy.collider;

                if let Ok(enlarged_aabb) = aabbs.get(entity) {
                    // Update the proxy's AABB.
                    update_proxy_fn(tree, proxy_id, Aabb::from(enlarged_aabb.get()));

                    // Record the moved proxy.
                    let proxy_key = ColliderTreeProxyKey::new(proxy_id, tree_type);
                    if moved_proxies.insert(proxy_key) {
                        tree.moved_proxies.push(proxy_id);
                    }
                }
            };

            // Clear the least significant set bit
            bits &= bits - 1;

and that stops the panic for me... I think it's like.. the proxy is set for the entity, and then the entity despawns and then it breaks here

[little maple]

I'm not sure if the above is necessarily a fix.. this prevents the panic and can get my game running, but it seems like collisions aren't always being detected 🤔

[dry quartz]

Hi guys

I have a crush on this line :

let proxy2 = tree.get_proxy(proxy_id2).unwrap();

on ..\avian3d-0.6.0\src\collision\broad_phase\bvh_broad_phase.rs

anyone have this issue?

the erreur is :

21:33:30 [windows] called `Option::unwrap()` on a `None` value                                                                   
21:33:30 [windows] stack backtrace:                                                                                              
21:33:30 [windows] note: Some details are omitted, run with `RUST_BACKTRACE=full` for a verbose backtrace.
21:33:30 [windows] Encountered a panic in system `avian3d::collision::broad_phase::bvh_broad_phase::collect_collision_pairs<()>`!
21:33:30 [windows] Encountered a panic in system `avian3d::schedule::run_physics_schedule`!                                      
21:33:30 [windows] Encountered a panic in system `bevy_app::main_schedule::FixedMain::run_fixed_main`!                           
21:33:30 [windows] Encountered a panic in system `bevy_time::fixed::run_fixed_main_schedule`!                                    
21:33:30 [windows] Encountered a panic in system `bevy_app::main_schedule::Main::run_main`!                                      
21:33:30 [dev] Application [windows] exited with error: exit code: 101```

[surreal junco]

Try updating to 0.6.1; there were a couple of crashes fixed

[hoary turret]

Is the ColliderDensity component always guaranteed to be present on a valid collider?

that is, if I spawn a collider as a child of a rigidbody, will it have the ColliderDensity component with a value of 1.0 inserted by default, or do I have to account for this and query for Option<ColliderDensity>?

[dry quartz]

I will but tomorrow 😄

[sweet sundial]

https://docs.rs/avian3d/0.6.1/src/avian3d/collision/collider/parry/mod.rs.html#363

[hoary turret]

woah, that's the most concise answer I've received in quite a long time

appreciate it hahaha

thanks

❤️

[sweet sundial]

o7

[dry quartz]

it's fine now no crush register, Thanks

[dusk walrus]

is there a way to get the entity from MoveAndSlide::intersections? It seems like it only returns physical information about the hit (normal, position, some other more niche but useful stuff) but not what actual entity it's currently talking about having found. It would be useful for the controller I'm writing to get both pieces of information from this API (I'm using it to check grounding)

[heady oasis]

by default does SpacialQuery.cast_ray() only include objects with a rigidbody and a collider or is a mere mesh + transform enough?

[sweet sundial]

definitely doesn't pay attention to Mesh

[dusk walrus]

Needs at least a collider, and there probably needs to be a rigidbody in either the same entity as the collider or an ancestor

[royal helm]

does shapecasting work with voxel colliders?

[surreal junco]

If I remember correctly it should only need a collider, since the cast apparently can also hit sensors (forgot to reply @heady oasis)

[vestal minnow]

yes but voxels don't support other non-convex or compound shapes, that's a limitation on Parry's side

[royal helm]

gotcha, that should be fine for me, just trying to debug an issue with my tank shapecasting not getting the correct points to push off of

that might be worth mentioning in the docs though

I remember seeing something about that in the rapier/parry release

[royal helm]

actually does not also not work against other voxel colliders too?

since they are non-convex

I'll test it

[vestal minnow]

I think that should work but not 100% sure

[royal helm]

main confusion to me is what these yellow spheres and arrows are doin over here lol

ok ya the difference between hit.point1/point2 are pretty big for some reason

odd

[livid star]

Is there a way to snapshot some of the physics environment and query it in background threads / bevy tasks? im aware the intersections and cast results would be stale but in my context thats acceptable.

[nimble fiber]

Usual disclaimer: I'm not an Avian expert just tossing some ideas out there

I imagine because Avian is designed to use the ECS "as much as possible," the way to do this is to query whatever entities you care about in a system on a timer and clone them off to a separate sub-app or similar structure, and do whatever computation you need to do on that subapp.

[livid star]

Yeah, but afaict id have to provide my own broad and narrow phase solution.

[nimble fiber]

I... Believe subapps have their own set of plugins

[livid star]

tbh i didnt consider a subapp, just jumped to bevy tasks.

[nimble fiber]

It might be a bit janky, I know the maintainers are still discussing a successor to subapps as a multiple worlds approach so documentation might be sparse

But yeah maybe a subapp with the Avian plugin loaded and you just clone over whatever entities you need to that subapp?

It's more manual work than like, Jolt's snapshot function where it'll just archive the world state and any deltas for you

Which is what I'm guessing you're looking for

[livid star]

yeah that sounds like what im after

[dusk walrus]

is there a better way to have a controller using MoveAndSlide collide with entities it bumps into? Currently I'm applying a slightly larger "real" collider to the entity than is used for the move & slide so that it will slightly clip in and collide

[vestal minnow]

the KCC examples in the repo have a basic implementation for it, it's basically the same as what bevy_ahoy does
https://github.com/avianphysics/avian/blob/25db6df55aaa1fa460b005d3ee2f341cbf10b918/crates/avian3d/examples/kinematic_character_3d/plugin.rs#L456

however it's certainly not perfect, the current version often makes tiny objects fly away

[dusk walrus]

to be clear i don't just want to apply forces, I mean for CollisionStart events to fire. It's fine if it doesn't inherently apply forces

[vestal minnow]

ah

no good way currently apart from manually triggering the events

or a larger "real" collider like you're doing

[royal helm]

definitely think there is a slight bug with the voxel collider shapecasting, I'll try to make a minimal repro at some point but it might be on the parry level

its completely work-aroundable, but my guess is the 2nd hit point is on a single voxel that isn't being translated to its actual position in the voxel collider

[proven prism]

Hi, I’m new to Avian and I’ve run into a couple of issues that might be bugs—or possibly just gaps in my understanding.

First, I noticed that when an entity does not have a Transform component, the Collider’s AABB seems to shrink to an extremely small size. This causes raycasting to behave incorrectly.

Second, when there is no Transform component, the PhysicsDebugPlugin also fails to render the collider. I’m not sure whether this is because the collider becomes too small to visualize, or if the debug rendering system explicitly depends on Transform.

I would also like to kindly request that Avian be updated to support the Bevy 0.19-dev branch. Since this branch has just merged BSN, I’m very excited about it and can’t wait to start using it.ort the Bevy 0.19-dev branch. Since this branch has just merged BSN, I’m very excited about it and can’t wait to start using it.

Thanks!

[surreal junco]

I believe Avian as a whole relies on the Transform component existing; it gets copied to the Avian-specific translation and rotation components, sure, but I believe Transform is still required to exist

Adding a Transform::default() should help; maybe Transform should be added as a required component?

[proven prism]

I understand that Avian already provides its own components for position, rotation, and so on, so I’m confused about why it still depends on Transform.

In my project, many entities are intentionally designed without Transform, which is how I ran into this issue. As a temporary workaround, I added Transform components to those entities, and it fixed the problem—but it feels unsatisfactory.

Directly modifying Avian’s Position and related components works fine. The only remaining issue now seems to be the dependency of scale on the Transform component.

[vestal minnow]

Definitely shouldn't depend on Transform for the core simulation, unless there has been some recent regression there

We have a bunch of tests that only use Position and Rotation and work fine

But I'll test once I get home

Scale and entity hierarchies might not fully work without transforms but otherwise it should be fine afaik

[proven prism]

Yes, that’s exactly the issue I ran into the scale doesn’t work.

[vestal minnow]

Yeah, simulation and debug rendering without Transform definitely works

just scale doesn't

[vestal minnow]

I would also like to kindly request that Avian be updated to support the Bevy 0.19-dev branch
There will be an Avian 0.7-rc release when Bevy's own 0.19 release candidate comes out (some time within a week probably?). I believe @untold peak has a local 0.19 branch for Avian already, which will probably be up fairly soon

[untold peak]

i am working on this now yes

[vestal minnow]

I intend to also experiment with some collider Template stuff once it's out

[untold peak]

@vestal minnow how do you handle bevy_heavy and bevy_transform_interpolation

for your version upgrades

[vestal minnow]

update them with a bevy version bump

and bump the 0.X.0 version of the crate

[proven prism]

I tested the following code,I just confirmed that when I remove the Transform component, my raycasting fails and nothing is drawn.

[acoustic spoke]

WITHIN A WEEK???

whoa, i really need to return to bevy again

[untold peak]

ye i've forked them as well

i'll get them all PRd shortly

just fyi, they will be tracking bevy main

as i needed to in order to get jackdaw to compile with the avian deps

but hopefully the breaking changes will all be resolved for avian, and you just need to change to 0.19.0 once it's cut

[proven prism]

Oh, my mistake. I customized a plugin and disabled some features I didn’t need. But I don’t think what I disabled would affect scale, right?"

I found the problem: I had disabled the PhysicsTransformPlugin because I didn’t have a Transform and didn’t need to sync data, and that’s when this issue occurred.

[untold peak]

should have a bunch of pr's coming now @vestal minnow

i've set it to track bevy main on the main branches of the subcrates, just shout if you would prefer to leave individual bevy tags

[vestal minnow]

I think I'll merge them once 0.19-rc is out if that's fine (assuming that happens relatively soon), in the meanwhile it's fine to track bevy main

[untold peak]

i could setup bevy-main branches on each if you like

ok yea that's fine, i can just point my local avian integration in jackdaw to my forks

[echo parcel]

When generating collision from mesh, is there a way to reduce resolution of the collision? I have a very dense mesh and I definitely doesn't need such dense collision from it

[spring moss]

You could downres the mesh first or you could try convex_decomposition_from_mesh_with_config, I’m not sure exactly how all the params work but resolution is probably a good start

Here’s an example of using convex decomposition to reduce hull complexity from the vhacd repo, they’ve got a description of all the params on their readme https://github.com/Unity-Technologies/VHACD#parameters

[echo parcel]

Does it work with open meshes? Like Planes, chunks, parts of the terrain etc?

[spring moss]

Not sure! You’ll have to try it. I think avian treats faces in open meshes as thin and double sided but I’m not at my laptop so I can’t check

[echo parcel]

Ok, thanks

[echo parcel]

Seems like it worked on the flat surfaces

But its not exactly what I need. These are the params I used:

let vertices: Vec<Vec3> = buffer
    .positions
    .iter()
    .copied()
    .map(Vec3::from_array)
    .collect();

let indices: Vec<[u32; 3]> = buffer
    .indices
    .chunks_exact(3)
    .map(|tri| [tri[0], tri[1], tri[2]])
    .collect();

let params = VhacdParameters {
    resolution: 32,
    max_convex_hulls: 8,
    plane_downsampling: 8,
    convex_hull_downsampling: 8,
    ..default()
};

let chunk_collider = Collider::convex_decomposition_with_config(vertices, indices, params);

[dusty shore]

is there some reason SpatialQueryFilter takes a LayerMask instead of CollisionLayers? it seems to implicitly compare with CollisionLayers::new(LayerMask::ALL, self.mask), i'm migrating from rapier and have invisible walls on a layer that excludes bullets (which works for projectiles but not hitscan/raycast)

[spring moss]

I think you might need to set fill_mode to surface only or flood-fill with detect cavities enabled, but I haven’t used that myself so I don’t know for sure

[echo parcel]

In both cases it still happens

[spring moss]

Hmm, I don’t know sorry. Does the mesh have disconnected parts? That might be related

[thin osprey]

I remember a while back using joints meant the game couldn't be used commercially (I think it was joints anyway)
I'm pretty certain this got changed in recent versions, but is 0.3 one of those versions?

[nimble fiber]

Aren't joints still XPBD-based? I know rigidbody integration was moved to an inertia-based implementation but I don't think joints were.

[vestal minnow]

Joints still use XPBD yes, the rest is impulse-based

In practice the risk is extremely small (many MIT implementations already exist, the authors of XPBD said they're not aware of restrictions there, and Jolt also uses XPBD for soft bodies, and Godot uses Jolt) but the patent does technically exist so it's just something to be aware of

we do still plan on switching joints to something else eventually

other things have just been a higher priority for now

[echo parcel]

How can I use ray casts with open colliders? I have a chunk, that is just a piece of a mesh, and so its open. My problem is, when I cast ray trough that collider, its being ignored.

I use trimesh_from_mesh to create a collider

let mesh = construct_mesh(&buffer);
let mesh_data = MeshData::consume_surface_nets_buffer(buffer);
let chunk_collider = Collider::trimesh_from_mesh(&mesh).unwrap();

And then just add it into the entity:

commands.entity(chunk_entity).insert((
    Mesh3d(mesh_handle),
    MeshMaterial3d(handle.clone()),
    chunk_collider,
    CollisionMargin(0.1),
    //DebugRender::collider(Color::srgb(0.0, 1.0, 0.0)),
    MeshSaveSize::from_mesh_data(&mesh_data),
));

When I visualised the collider, it looked fine.
I use just simple ray cast:

    if let Some(hit) = spatial_query.cast_ray(
        origin,
        direction,
        max_distance,
        solid,
        &filter,
    ) {

But it returns none when pointed at the chunk

[hoary turret]

I find it really nice how the mass of the rigidbody gets distributed after changing the colliders' and their mass distirbutions

and it topples and loses its balance without applying any force

😃

[echo parcel]

How does it work? Is it procedural or does it have pre set breaking points?

[hoary turret]

well calling it "procedural" would be a huge overassesment

it's just a voronoi diagram

with some randomness

[frail robin]

In my game I only need ray and shape casting. Would it be possible to feature-gate other things like transform interpolation?

[somber portal]

Why might my avian3d colliders (walls) mostly only work on one side?

[somber portal]

I even enabled SweptCcd

[sweet sundial]

which side? what shape are they?

[somber portal]

I'll get a video later. But it's just a normal wall. I'm not sure if there are any specific sides, as I'm talking about a bunch of them with similar effects.

From one side I can walk through the wall. And from the other I get pushed on top of it.

I forget. Is there a debug mode to view the colliders?

[royal helm]

Are they cuboid colliders or convex hulls or trimeshes or?

[royal helm]

PhysicsDebugPlugin should add them

[somber portal]

Currently ConvexHullFromMesh.

If it helps, i'm using skein, so I'm attaching a colliderconstructor to a mesh in Blender.

These are the assets I'm trying to apply colliders to:
https://loafbrr.itch.io/interiors-a

I'm wondering if the walls are too thin

I suppose I can somehow create a cuboid collider to encompass both sides of the wall. So the collider would encompass the entire wall "space".

[echo parcel]

I have simialr issue, but my problem is that they don't work at all

[somber portal]

Here's a video showcasing my collider problem.

If anyone has any idea what's causing this, please let me know.

[somber portal]

Maybe it's something with my player collider?

This is weird; i'm even walking through cuboid colliders.

[somber portal]

Just made sure i'm using the latest version. Still no difference.

[somber portal]

Solved it. I needed to add static rigid bodies

[echo parcel]

I had "ColliderDisabled", dang...

[echo parcel]

Doesn't avian have any automatic way to determine whenever collision should be enabled or disable, or should I manually manage that?

[nimble fiber]

...under what conditions?

usually enabling/disabling collisions depends heavily on the game

[echo parcel]

I ask cause if I remember well engines like Unreal can determine collisions by themselves, based on e.g. distance. Just want to know if am I doing anything wrong, or is it by design

[silk mauve]

Like if you get far away from a physics object it "turns off"? Is that what you're thinking?

[echo parcel]

As far as I know, unreal uses octree based on distance to determine which colliders should be able to interact with each other to minimize amount of computations

[nimble fiber]

oh, yeah Avian does that

it's called Broad Phase

you don't really have to meaningfully worry about it. ColliderDisabled is mainly for like, game mechanic purposes, i.e. turning off a player's hitbox during iframes.

[echo parcel]

Btw. If I have 2 Colliders that touch each other, does they generate collision events and constantly check each other? I have tightly packed chunks, with colliders touching each other on the edges. Should I do something about that? Might that cause lag I experience? Or the mere presence of not-disabled collisions is causing that?

Should I somehow set that colliders of entities that use ChunkTag ignore other colliders of entities that use ChunkTag?

[vestal minnow]

static bodies don't check for collisions between each other, so it's not a problem if they overlap

if they're not both static then they will check for collisions and have some overhead, yeah

(on Avian 0.6)

[echo parcel]

Time to update and test, then

[echo parcel]

Seems like updating to 0.6 fixed my issue, thanks for the tip

[echo parcel]

When can trimesh_from_mesh return None? I guess when the mesh is empty but are there other examples?

[sweet sundial]

https://docs.rs/avian3d/0.6.1/src/avian3d/collision/collider/parry/mod.rs.html#1499 https://docs.rs/parry3d/0.26.0/parry3d/shape/enum.TopologyError.html

non-float points, zero length edges, weird winding, etc

[minor sandal]

Hi there, I'm trying to modify the transform of a child item of a physics entity, basically trying to make an icon stay the same angle regardless of object angle.

When I get very fast rotating object I still get an angle on the object's map icon as shown (slightly slanted and flickering X)

So far i've tried a whole bunch of combinations of schedules for my system so that it runs after the interpolation, but before the transform propagation... And I have not been able to do so...

Would any one have an idea of where in the scheadule i'm supposed to insert my system?

Any help would be extremely appreciated !

Here are are a few tries i did:

  app.add_systems(
      PostUpdate,
      update_rotation
          .before(TransformSystems::Propagate)
          .after(PhysicsSystems::Last),
  );

app.add_systems(
    PostUpdate,
    update_rotation
        .after(TransformSystems::Propagate)
);

app.add_systems(Last, update_rotation);

[sweet sundial]

https://docs.rs/avian2d/latest/avian2d/dynamics/rigid_body/struct.LockedAxes.html#associatedconstant.ROTATION_LOCKED

but for that kind of thing i usually just make a new relationship that only syncs translation

could modify just the GlobalTransform instead i suppose? either way you'd be doing and immediately undoing work

[minor sandal]

Ah, I'll have to look into relationships then most likely.

Thank you so much for your answer!

[thin osprey]

if a dynamic rigidbody collides with a static rigidbody, will it trigger a Change<Transform> on the static transform
I know it won't actually change the values, but does it cause a trigger
because I just found a dormant bug in my code and I think it's caused by this, I hope at least

[vestal minnow]

It shouldn't, but I'd have to test to make sure

[livid star]

is there an example showing avian physics and bigspace playing nice together? a quick discord/git search suggests there isn't.

[nimble fiber]

Check in with the big space thread in #1034543904478998539, the author was definitely showing off some physics engine stuff earlier

Idk if it's avian but

[vestal minnow]

it's rapier

[nimble fiber]

ah, my bad

[minor sandal]

I worked a bit on the child until I realised that it doesn't matter, I still need to insert myself at the correct spot in the scheadule, otherwise my icon is delayed if the unit is fast. Or if the icon should rotate, the icon will rotate late.

But at least I have the child working xD

So for now i'm still trying to insert myself before the global transform update, but after the physics interpolation update.

[sweet sundial]

physics/interp shouldn't be anywhere after fixed

[minor sandal]

Odd...
I've been trying to get it but the debug gizmo is still ahead of my sprite

app.add_systems(
            PostUpdate,
            update_icon_position.before(TransformSystems::Propagate),
        );

Nevermind.... It's fixed by doing it in last instead of postupdate!

Thank you so much for all the help!

[white garnet]

what would be the easiest way to create a "inverse capsule" collider? E.g. a sort of arena where you can be inside of

[sweet sundial]

several halfspaces

[white garnet]

is there any way to debug why PhysicsDebugPlugin isn't rendering anything? The physics work but the debug renders never show, even though I have debug-plugin in the features of avian2d. I'm wondering if bevy itself needs some feature I didn't enable or something

[vestal minnow]

do you have bevy_gizmos_render for Bevy?

[white garnet]

That was it thanks! is this missing in the docs ? Would you support a PR with log::warn! in case it's not there when the plugin is being built?

[frail robin]

The documentation says CcdPlugin is included PhysicsPlugins, but I don't see it there 🤔

It's missing in the table in the docs and I don't see it in the PhysicsPlugins group in the code.

[vestal minnow]

it's in SolverPlugins which is in PhysicsPlugins

it's documented for that
https://docs.rs/avian3d/latest/avian3d/dynamics/solver/struct.SolverPlugins.html

[vestal minnow]

Hmm, I believe one of the reasons bevy_gizmo_render is separate from bevy_gizmo in Bevy is to allow alternate rendering backends. I think it'd be strange for Avian to enforce a specific renderer when it doesn't really care about that, it just needs a gizmo API to call (which is provided by bevy_gizmo)

but we could add a mention of it in docs somewhere

[frail robin]

Ah, missed it, thank you 🙂

In my game I only need ray and shape casting. Would it be possible to feature-gate other things like transform interpolation?

[summer acorn]

if you only need ray and shape casting then thats just parry isnt it?

[nimble fiber]

i think i've seen that brought up before, and the problem is parry doesn't have a very bevy-friendly API

you need to write lots of glue code to get the two to play nice

[frail robin]

Exactly!
That's why I asking if Jondolf would be open to something like this or the plan is to use Peck for cases like this in the future.

[molten mason]

yay!
I solved my spring jiggle by moving the system from Update into PhysicsSchedule with .before(PhysicsStepSystems::First), but I wonder if there's a friendlier way?

[past cargo]

You should do Physics stuff on FixedUpdate

Update will have jiggle

[molten mason]

oh so just FixedUpdate is good enough? I'll try 😄

thanks 😄

[molten mason]

I have another problem with my grabbing mechanism. The spring works fine for linear constrained things, but with a RevoluteJoint it seems to do really nasty stuff with the velocity_at_point method.

My code looks like the following every fixed update:

// mut target: Single<(Entity, &GlobalTransform, Forces, &CursorTarget)>,
// mut cursor: Single<(Entity, &mut Position)>,

    if let Some((_, ray)) = rays.iter().next() {
        cursor.1.0 = ray.origin + ray.direction * cursor.2.depth;
    }

    let anchor = target.3.anchor;
    let want = cursor.1.0;
    let am = target
        .1
        .mul_transform(Transform::from_translation(anchor))
        .translation();
    gizmos.sphere(
        Isometry3d::from_translation(want),
        0.01,
        Color::Srgba(Srgba::GREEN),
    );
    gizmos.sphere(
        Isometry3d::from_translation(am),
        0.01,
        Color::Srgba(Srgba::RED),
    );
    let force = (want - am) * 20.0 - (target.2.velocity_at_point(am) * 5.0);
    gizmos.arrow(am, am + force, Color::Srgba(Srgba::BLUE));
    println!("{force}");
    target.2.apply_force_at_point(force, am);

and the further away I am from the joint, the worse it gets

anyone know what I can do to fix this?

[viral goblet]

Would you say 0.6.1 is more stable than the latest 0.5?
I'm close to a steam demo and people are still having crashes due to parry

(But in the last two times I tried migrating to an 0.6 the game crashed in the first minute or so)

[sweet sundial]

you can tell bevy to just log and keep going if a system errors

[nimble fiber]

what is actually causing the crashes? parry is decently mature for the rust world, maybe not as much as something like Jolt or Box2d but regular crashes shouldn't be common

[silk mauve]

I had a parry crash on 0.5 and upgrading to 0.6.1 fixed it for me

It was also really straightforward. Literally had to change 0 code

I think it was this one, or at least similar:
#1124043933886976171 message

[vestal minnow]

yeah that crash was because of Parry's BVH, which we no longer use 😄 (we switched to obvhs)

[viral goblet]

Wait you can? There's a way to catch panics always? 😮

[viral goblet]

It doesn't happen very often, but once every 20 sessions when you have 2k play sessions is a hundred crashes

[viral goblet]

I already changed all &mut CollisionLayers to try_insert so I'm ready to move once stable

[vestal minnow]

nah that's for systems that return Results, or for commands that fail, not for actual panics or unwraps

Rust does have catch_unwind, but it's not a full panic recovery thing, mostly useful for e.g. saving progress or logs to a file before quitting for example

[viral goblet]

I see

I already am printing the crash log to a file so release builds can share them after crashes

As head of Avian, would you say 0.6.1 is more stable?

[vestal minnow]

It's faster and fixes some crashes (like the BVH crash Big Yolk was seeing before), but also potentially introduces a few new obscure ones

I didn't see any Avian crashes in the Bevy jam game we did with 0.6, but it depends on the game and what weird things one might be doing

[sweet sundial]

https://docs.rs/bevy_ecs/0.18.1/src/bevy_ecs/schedule/executor/single_threaded.rs.html#157

(correct but i wanted to get details)

[vestal minnow]

notably entity cloning has never been fully supported, and entity disabling is supported but is fairly bug-prone in general (not just in Avian but in all of Bevy)

[sweet sundial]

-# (also technically every panic unwind goes to a catch_unwind, there's one right around main() that handles actually quitting the process on panic)

[molten mason]

Like I don’t understand what happens here. When I grab further away and just slightly move my mouse the arrow is already pretty large and makes the door spin weirdly sideways. But shouldn’t the torque that’s being added because of the further distance do the opposite? Or is there some weird edge case that I need to manually add torque on velocity_at_point?

[sweet sundial]

what happens if you turn the force down

[molten mason]

you mean the * 20.0?

or just dragging down?

I made the * 20.0 a * 0.2 but it still multiplies till it gets too big and crashes

tho normal dragging for linear things is unusable

huuuh

I don't even have to move my cursor

somethings definitely off with the apply_force_at_point to velocity_at_point thing

okay without moving my mouse, I currently do this:

    let anchor = target.3.anchor;
    let want = cursor.1.0;
    let am = target
        .1
        .mul_transform(Transform::from_translation(anchor))
        .translation();
    gizmos.sphere(
        Isometry3d::from_translation(want),
        0.01,
        Color::Srgba(Srgba::GREEN),
    );
    gizmos.sphere(
        Isometry3d::from_translation(am),
        0.01,
        Color::Srgba(Srgba::RED),
    );
    let force = (want - am) * 0.2 - (target.2.velocity_at_point(am) * 5.0);
    gizmos.arrow(am, am + force, Color::Srgba(Srgba::BLUE));
    target.2.apply_force_at_point(force, am);
    println!("{force:.8} -> {:.8}", target.2.velocity_at_point(am));

and get the following result

with a damping of 0 or with a damping of 100 it both gets worse

why does it do such weird feedback for revolute joints 😭

[viral goblet]

Ok so I updated to 0.6.1 and will be now reverting back to 0.5
After about 30 seconds of gameplay I get:

no element at index 1 in StableVec::index_mut

Encountered a panic in system `avian2d::collider_tree::update::update_moved_collider_aabbs<avian2d::collision::collider::parry::Collider>`!
Encountered a panic in system `avian2d::schedule::run_physics_schedule`!
Encountered a panic in system `bevy_app::main_schedule::FixedMain::run_fixed_main`!
Encountered a panic in system `bevy_time::fixed::run_fixed_main_schedule`!
Encountered a panic in system `bevy_app::main_schedule::Main::run_main`!```

[bleak wadi]

is it possible to use Position rather than translation? this might be a tnua issue. also maybe an xy problem. what i'm trying to do is keep track of a very precise position value, but keep the translation locked to a large pixel grid

if i round() the translation value on FixedUpdate, there is a strange shimmering. if i do the same on Update, the movement only updates every few frames for some reason, but is not shimmery

it's kinda jerky

i'll record a video to show the different outcomes

[sweet sundial]

can replace the sync systems

PhysicsPlugins::default().build().disable::<PhysicsTransformPlugin>()

[bleak wadi]

alright, it looks good now, but collision seems to be disabled

[bleak wadi]

does anyone know why that is?

[sweet sundial]

are you initializing position and rotation?
try enabling the debug plugin

[echo parcel]

How can I remove impact force? When I hit a wall it causes player(With sphere collision) to rotate. I want it to just stop, without any turning, rotating, impact forces etc.

[nimble fiber]

Kinematic body instead of dynamic?

[echo parcel]

When I used Kinematic Body collisions doesn't work. I also still want to use forces like e.g. Gravity,

[sweet sundial]

fixedaxes?

[echo parcel]

Whats that?

[sweet sundial]

locked*
https://docs.rs/avian3d/latest/avian3d/dynamics/rigid_body/struct.LockedAxes.html

[echo parcel]

Got the desired effect with:

        RigidBody::Dynamic,
        LockedAxes::ROTATION_LOCKED,
        Restitution::new(0.0),
        Friction::new(1.0),

Thanks

[livid wolf]

Hello everyone. Does anybody know how can I get "the actual velocity" of a body in avian2d?
For example if i have a body that has a non-zero velocity but is stuck at a wall, i want to know that "the actual velocity" is zero

[nimble fiber]

Dumb answer: Make a system and component which just tracks changes to Transform frame over frame and calculates velocity from that.

Maybe smarter answer: ContactPoint has a field which returns the delta velocity of a collision, which I think would be useful with this.

[sleek thicket]

isn't linear velocity already the actual velocity? why do you need that in the first place?

[livid wolf]

Now that you asked i doubt myself 😃. I am not at pc so I will double check tomorrow, but i think it's not actual velocity.

I need it to tie walking animation speed to it. So that when my guy is walking into the wall he doesn't move legs.

[echo parcel]

How often does Avian update physics? During Update, FixedUpdate, some other thing? I want to cast shape every physics update

[echo parcel]

But, from what I see shape_cast returns only 1 hit. I want to make a character that sticks to the ground, basically a spider, so having more points would help to stabilize it. How should I approach that?

[sleek thicket]

you'll need raycasts for IK anyway, so probably that
and it's in fixed(post)update

[nimble fiber]

note: do not put your systems in FixedPostUpdate, learned that the hard way. Avian does not like it when you delete entities while it is updating physics.

[nimble fiber]

call shape_cast multiple times

[echo parcel]

So put it in the fixed update, right?

[livid wolf]

confirmed - linear velocity is not "actual velocity", even if the object is stuck it will still show some velocity. At least in my case, as I apply it manually to move the character

[past cargo]

No, LinearVelocity is a Way to apply velocity

It need Have a AssemblyLinearVelocity to know the actual

[indigo mango]

Hi folks !
is there any way to automatically create a Collider on the root entity , when using ColliderConstructorHierarchy with gltf scenes ? Thanks !
ie

    commands.spawn((
        Name::new("Test"),
        SceneRoot(
            asset_server.load(GltfAssetLabel::Scene(0).from_asset("models/power_cable_coil.glb")),
        ),
        ColliderConstructorHierarchy {
            default_constructor: Some(ColliderConstructor::ConvexHullFromMesh),
            ..Default::default()
        },
        // automatic root collider ?
        RigidBody::Dynamic,
        Mass(20.0),
    ));

[vestal minnow]

The colliders are on the entities that have the meshes, which for glTFs are always child entities afaik

You could implement some custom thing that collects all meshes and creates a single compound collider, but that's not a built-in feature currently

[frail robin]

I implemented what Jondolf mentioned for my game, feel free to use it:
https://github.com/simgine/simgine/blob/project_harmonia/base/src/combined_scene_collider.rs

[sleek thicket]

can't you schedule your check before you modify LV?

[past cargo]

if he modify the vector velocity

it wont show on LV

[indigo mango]

Ok thanks a lot for the information ! Btw, Avian is absolutely fantastic, the features , performance & API is a joy !

[indigo mango]

Wow, very nice , thanks a lot ! I had started on implementing it, but this is cleaner , thanks 🙂

[silk mauve]

When I do this:

commands.entity(e).insert((
   RigidBody::Dynamic,
   Transform::from_translation(position)
));

the entity will always go to (0,0,0) in the world instead of my position. It seems like RigidBody::Dynamic inserting Position::PLACEHOLDER is overwriting the Transform that I inserted? On further testing this works fine for example:

commands.entity(e).insert((
   RigidBody::Dynamic,
   Transform::from_translation(position)
));

// just re-inserting transform AFTER RigidBody
commands.entity(e).insert(Transform::from_translation(position);

Is this intended behavior? I ended up doing this:

commands.entity(e).insert((
   RigidBody::Dynamic,
   Position::new(position),
   Transform::from_translation(position)
));

[quick sluice]

Hi, I am currently creating a game where a player with a rectangle (or round_rectangle) collider is walking on a 2d tilemap. each stone has it’s own static collider. I move the player by changing it’s velocity. The player still walks into walls when it collides with two blocks standing next to each other and only gets pushed out slowly. How can I fix that?

Thanks for your help!

[molten mason]

I currently apply a big force one time at a rigidbody which is axis locked and on a prismatic joint. Currently, when it's at the end, it just stops. How can I make it "rebounce" a bit easily?

[hoary turret]

Yo @vestal minnow I think I found an issue, but I'm not sure whether it's intended or not, so asking here b4 opening an issue/pr in github
Resizing the collider/s of an sleeping rigidbody doesn't wake it up (Haven't checked but changing the transform overall probably doesn't)
Which basically puts the rigidbody in a physically inaccurate position until it's waken up somehow and re-checked.

Is this intended behavior? or would you accept a PR that makes transform propagation wake up the body if it's sleeping?

[vestal minnow]

Changing the translation or rotation does wake it up, but changing scale doesn't currently. I'd accept a PR to allow changing scale to wake up entities

as a workaround in user code, you can just .try_remove::<Sleeping>() when changing scale, or alternatively use commands.queue(WakeBody(entity))

[vestal minnow]

There's a few physics engines that have a joint limit restitution for this (PhysX comes to mind), but we don't have that built-in currently. You could make the limit itself soft by changing its compliance, but that'll just make it springy and not "bounce" against a hard limit

you could also manually try to detect when it hits the limit, compute the velocity at that point, and then apply some rebound impulse to emulate the effect

just a bit more complicated

[vestal minnow]

Look up "ghost collisions" in this Discord or in the repo. But for this case:

• If you used Collider::voxels for the tilemap, it wouldn't have the problem (except maybe at the edges between different colliders...)
• Adding SpeculativeMargin::ZERO should also make it less of a problem, with the caveat that your character could technically clip into or through walls at high enough speeds, since it disables speculative collision
• There are some more specialized solutions that you could try, like the approach in this article

[quick sluice]

Thank you :) I just tried collision compound and it is working now

[vestal minnow]

Not intended behavior, but I'm not too surprised it's broken there 🤔 the transform initialization logic is very scuffed haha

would be good to try making it more robust though

[silk mauve]

Had me going for like 3 hours lol. Makes sense though

[hoary turret]

yeah, that's what I'm doing

[echo parcel]

When I'm setting movement direction I'm modifying LinearVelocity, but then I'm overwriting Velocity set by ConstantLocalLinearAcceleration. How should I resolve this? I suppose creating my own component for things like gravity isn't the ideal solution, since ConstantLocalLinearAcceleration exists for a reason, right? I also can't just leave Y axis for the gravity, and modify other axes since gravity in my project can be from any direction

The other question is, how can I reset acceleration? It seems like its being added indefenitely:

/// Applies [`ConstantLocalLinearAcceleration`] to the accumulated local acceleration.
fn apply_constant_local_linear_acceleration(
    mut bodies: Query<(
        &mut AccumulatedLocalAcceleration,
        &ConstantLocalLinearAcceleration,
    )>,
) {
    bodies
        .iter_mut()
        .for_each(|(mut acceleration, constant_acceleration)| {
            acceleration.linear += constant_acceleration.0;
        })
}

But clear_accumulated_local_acceleration is private.

[glossy holly]

what counts as a "very high subset count"?

0.6 seems to have breaking CollisionLayers changes

it's been changed to an immutable component i see

[vale dragon]

hey there! I have an Avian plugin crate that is getting ready for a 0.1.0 release, is there any convention on crate names? I was thinking on going with avian_<specific-plugin>, but I don't want to ruff feathers

[glossy holly]

is it possible to enable/disable debug gizmos in runtime without having to manually do for each entity

[nimble fiber]

Yeah I have a checkbox that does this, gimme like 10 minutes to wake up and I'll send it to you

[glossy holly]

tyty

[nimble fiber]

fn toggle_physics_debug_vis(gizmo_config_store: ResMut<GizmoConfigStore>) {
  gizmo_config_store.config_mut::<PhysicsGizmos>().0.enabled
}

set that to alse to turn them off and true to turn them on

[glossy holly]

yipee

ty

yup worked perfect ly

[sleek thicket]

nope, people just call their crates whatever they want

[vestal minnow]

avian_* is fairly common, see for example avian_pickup or avian_rerecast . If it's something that a physics engine would not be expected to have built-in, go for it! But if it's a more general name like avian_solver or avian_collision, or you're otherwise unsure if it's okay to claim, it's good to ask first

There's no specific naming scheme though, some might have a bevy_X crate and the Avian integration is bevy_X_avian, or some crates might use a fully unrelated name, like bevy_ahoy

[vestal minnow]

Honestly that sentence might be kinda outdated or inaccurate, I think if you have so many substeps where numerical issues actually become a problem, your performance would probably be dead anyway

I don't have any specific numbers there

[glossy holly]

i ended up increasing the fixed schedule rate for my usecase

[vestal minnow]

mostly just tune it such that stability is good enough while things still run well

[glossy holly]

and everything seems fine

[vestal minnow]

yeah that works too, the effect is slightly different but both can improve stability

256 Hz feels a bit on the high end, but if it works well, then that's fine

[glossy holly]

had issues with ghost collisions

[vestal minnow]

yeah that makes sense

[glossy holly]

if a high speed object moved towards a wall, and a spike was behind that wall, it would hit both the wall and the spike

so i increased the physics rate to x4, which still caused the ghost collision

so i required the object to collide with the spike for 4 frames in a row

and that finally fixed it

which is why i did fixed update rate instead of substepping

[vestal minnow]

there's SweptCcd that you could use instead of speculative collision (or alongside it) that likely wouldn't have that issue, although it has its own caveats

[glossy holly]

i did use sweptccd

[vestal minnow]

hmm okay

[glossy holly]

sweptccd is on the entire process

im curious why the ball collides with the spike for multiple frames in a row though

when there is a ghost collision, why isn't it just a one frame ghost collision?

[vestal minnow]

not sure, hard to tell without running the scene

I would intuitively expect it to be at most two timesteps, ideally one

[glossy holly]

same but I don't have much understanding of physics simulation

hmm

seems to be 2 frames actually

2 frames of ghost collision

even without the absurd high fix timestep rate setting a 3 frame requirement seems to fix it

yippee slight improvement

[vale dragon]

thanks! I think it's fine then, I don't expect avian to have it built-in 🙂

[glossy holly]

I want to use a param set within collision hooks but since it's not mutable access i'm not able to

i want to create multiple different checks in a row, and don't want to deal with query conflicts

is there a way to make this work, or should i just use a few with<> and without<>s

what's the difference between these two?

[raven oxide]

One returns a vector (direction), one returns a magnitude / number (length)

[vestal minnow]

Pretty much what the doc comment says. If you have two opposing impulses (e.g. an object wedged between two surfaces), then they could cancel out and the total impulse vector would be near zero; but the scalar impulse magnitude would be the sum of the magnitudes of the individual impulses, disregarding direction, making it non-zero

[glossy holly]

Oh I see

interesting

when would the vector be useful?

[vestal minnow]

The vector is useful for "what is the final impulse acting on this body" and the scalar magnitude is useful for "how much impulses are being exerted on this body in total" (could be used for some threshold for breaking the object into pieces under high pressure)

[glossy holly]

Yup, the latter is what I ended up implementing, and magnitude worked perfectly for that

[vestal minnow]

If you really need to use mutable logic, you can queue a command with commands.queue(...), which has exclusive world access

Or run a one-shot-system

With commands.run_system(...) IIRC

which would let you easily use system parameters

there's also something like run_system_with if you need to pass input

The caveat is that you can't modify the contact there directly anymore, and it's run deferred

(since it's a command)

[glossy holly]

I see, well the goal would be to modify the contact

but that seems really useful for code organization in general

[vestal minnow]

The reason mutable access isn't allowed in the hook is because it's run inside a parallel loop where it's computing contacts, and so it can't provide safe mutable access to the hooks from multiple threads simultaneously

the provided commands are the main "escape hatch" to still allow safe deferred mutation that is run outside the parallel loop

I guess we could mayyybe also provide some unsafe variant that does allow mutable access to some capacity, but I haven't looked at how viable that is, seems sketchy at best

[glossy holly]

that makes sense yeah

the difficulty for me is that you can only have one set of collision hooks

and i have multiple different checks and effects I'm implementing

which can cause conflicting queries

even though the queries are done separately

[vestal minnow]

Yeahh I think you need With and Without filters then unfortunately, if ParamSet requires mutable access :/

it feels kinda weird if it requires that though

[glossy holly]

yup makes sense, tyty

[glossy holly]

finally found a decent way to check which collider is which component

what I was doing before

should work great for handling collision events in non-observer scenarios

maybe could be a macro? it's a common pattern I encounter

[glossy holly]

setting a rigidbody through transform seems to cause

tiny tiny floating point differences in position and rotation

:(

so the same transform can seemingly have slightly different rigidbody situations

unsure if this makes sense, but that's the behavior I'm seeing

will try setting position and rotation directly maybe,

or write my own syncing system?

[echo parcel]

Is there a way to change LinearVelocity but in local space?

[echo parcel]

Or maybe a way to apply force just for 1 tick, without accumulating/accelerating. So if this is just a single force and its removed then in the next tick entity just stops. Doesn't decelerate, just stops

[nimble fiber]

what are you actually doing in your game? Is this a momentum-preserving dash?

[echo parcel]

I want to create a movement system where gravity is not working in just 1 direction. Basically, its perpendicular to the wall character is standing on. I've seen examples where movement was done via setting LinearVelocity component, but this unfortunately resets ConstantLocalLinearAcceleration.

I'm simply trying to create 3 axis movement system

[nimble fiber]

I think you're looking for the Forces QueryData: https://docs.rs/avian2d/latest/avian2d/dynamics/rigid_body/forces/index.html#one-time-forces-and-impulses

I think the Constant* components will be less ergonomic because you'll be changing them pretty much every frame anyways as you only get one per entity

[echo parcel]

I tried using ConstantForce, but it accelerating, and decelerating

[nimble fiber]

yeah I don't think they're going to help you very much

you should be using Forces::apply_force

it's a per-frame force application

i'm assuming you have like, "gravity areas" or some such which apply gravity to an entity

you'd essentially have a sensor that checks for overlap with your gravity areas with your entities which you are applying gravity to, then calculate apply_force for each one based on the sum of forces desired from all overlapping areas.

[echo parcel]

Yes, basically every character is checking if its on the wall, if not, it applies gravity on itself

[nimble fiber]

yeah try Forces. The intention of the various Constant* components is, I suspect, just ease of use. If you want to get that granular you're gonna want to use the lower-level API.

[echo parcel]

Using apply_force works like using ConstantForce. So its accelerating, and then decelerating

[nimble fiber]

What does this mean exactly? Can you describe in more detail how the rigidbodies are behaving in a way you don't want?

[sleek thicket]

mario galaxy?

are you adding or setting LV?

[echo parcel]

When I press a button instead of going at the given speed and then immediately stopping when button is released, it’s gaining speed over time and then loses it over time when button is released

[echo parcel]

Ye, something like that:
https://youtu.be/iFAT6BqhE5A?t=29830

I've tried both setting and adding LV(Linear Velocity), but both has problems. Setting it resets other velocities like ConstantLocalLinearAcceleration, and adding it causes accumulation

[nimble fiber]

So the tl;Dr is that Super Mario Galaxy has cartoon physics and you can't achieve that by just setting acceleration or LV on a rigidbody. You need a character controller, not realistic rigidbody simulation.

You can write your own character controller from scratch but this is pretty nontrivial. Doing it with a kinematic body is easier because you set velocity and call a move-and-slide function, but then you often have to handle stairstepping and other little details. Rigidbodies are harder as you've encountered because you have to carefully calculate acceleration to achieve desired velocities.

You'll want to check out bevy_tnua and bevy_ahoy at least to read the code, if not to try them out and seeing if they will support Mario Galaxy style movement.

bevy_tnua also links to some good reference videos on YT that explain how it was made, which may help you.

You... might be able to achieve it by setting LV directly which is actually a decently unusual property of Avian

But not by accumulating it

[echo parcel]

I don't meen cartoon physics, just sticking to the walls, like on this video fragment I found. Can also be like this scene from Berserk, or this mod from minecraft:
https://www.youtube.com/shorts/wtCgQ9JIYdw?feature=share

I have all the stuff required for rotating player according to the fall and stick it, so I sort of have the controller, just need a way to move the player

[nimble fiber]

Can you send a video/gif of your character controller in-engine with one of your existing solutions? It'd help to see what exactly is going wrong.

I maintain you need some kind of character controller logic to handle your acceleration in order to cleanly limit your character to a maximum velocity.

For example this blog post goes over the Source engine's player character controller with some explanation: https://adrianb.io/2015/02/14/bunnyhop.html

[echo parcel]

This is version using forces.apply_force, and it feels like moving on ice

for (_, player_transform, _player, mut forces) in player_query.iter_mut() {
        let mut fly_direction = input.movement.x * player_transform.forward();
        fly_direction += input.movement.y * player_transform.right();

        let mut normal_dir = Vec3::ZERO;

        if let Some(normalized_direction) = fly_direction.try_normalize() {
            normal_dir = normalized_direction;
        }

        //linear_velocity.0 = normal_dir * 8.;
        forces.apply_force(normal_dir * 8.);

[nimble fiber]

Yeah read the article I sent you :P and do some digging into what a character controller is. The tnua and ahoy repos are good places to start as well.

Physics engines don't really consider it their responsibility to make characters feel good, and characters running on raw acceleration or directly setting velocity to constant values don't feel very good, as you described "sliding on ice." You need to introduce additional dynamics like friction, max velocity capping and altering acceleration depending on current speed.

[echo parcel]

Welp, I guess its time to learn then, thanks for the calrifications

[sleek thicket]

i did that with the very very valet controller, everything was mostly fine except camera, i couldn't figure out how to make it not nauseating in first person

[sleek thicket]

you just need directional damping

[echo parcel]

I see there is LinearDamping, but I don't see anything related to directional damping

[sleek thicket]

DIY 😅

[echo parcel]

Forking time? xD

[sleek thicket]

you can use linear, you'll just have to account for unwanted damping on the gravity axis too

[echo parcel]

Well, after setting gravity to the level of Jupiter's, I think it works fine with damping

[sleek thicket]

you can do more fun stuff if you make a directional one anyway, check very very valet's car video

[echo parcel]

This one?
https://youtu.be/CdPYlj5uZeI

[viral goblet]

I love this crate

[frail robin]

In my game I support pause. When the game is paused, time dilation is set to 0, everything freezes. FixedUpdate doesn't run.
But users can move objects during pause.
To pick objects I use avian and I'd like to keep it functional during pause. Just for testing I set the Avian schedule to PostUpdate. But it still breaks. On the video you can see that I can move objects freely while the game is unpaused. But once I pause the game, I can move any object only once. After that hovering over the object does nothing (you can see white outline when the object is hovered).

I tried using regular ray-casting instead of picking and it has the same issue - it breaks when the time is frozen 😢

[prisma raft]

Hello. I have problems with a "car simulation" using Avian3d.
I have pretty simple setup with main car body, wheels connected to it using RevoluteJoint. Car behaves extremely bad, bumps even on a slightly varying polygonal surfaces and get stuck into the ground quite fast. Where should I look to solve this problem?

I use Ccd, all bodies have masses, car has angular inertia

https://gist.github.com/BlazarBlade/395132e8eccc92da410f55d9a838c329

[frail robin]

I forgot to convert my video into Discord-friendly format and unfortunately Discord doesn't allow edit attachments. So here is the video in a new message.

[vestal minnow]

Can you try adding the update_moved_collider_aabbs system to run while the game is paused?

that should in theory update the spatial query BVH when you've moved colliders around

we should probably expose a more user-facing system or API for it

[frail robin]

I tried this:

        .add_systems(PostUpdate, update_moved_collider_aabbs::<Collider>)

But the issue is still present 🤔

[vestal minnow]

heck

hmm

[frail robin]

It was with manual shape casting. Let me try picking. I could be a multisystem issue 🤔

No, same problem.

@vestal minnow wait, no, picking works.

RayCast still breaks.

@vestal minnow
Sorry, just to clarify - clicking on colliders works via picking works. After I apply movement (done via transform manipulation and unrelated to picking and avian) - the collider is correctly updated when I add the suggested update_moved_collider_aabbs.
But the mentioned RayCast component is still broken. Its global direction always points to [1, 0, 0], no matter how I update its local direction.

[vestal minnow]

oh yea RayCaster and ShapeCaster are separate from that stuff 🤔 I suppose we could probably move their update logic out of the PhysicsSchedule so that they'd be updated even when the simulation is paused

[frail robin]

That would be nice.
Do you want me to open an issue?

[vestal minnow]

Yup, that'd be nice, thanks!

[vestal minnow]

Hmm does the SweptCcd actually help here or is it better if you remove it?

I feel like it might be some pathological case where it's detecting that it should stop at each of the edges which makes movement very stuttery/bumpy, and more generally you're probably seeing ghost collisions (you can search this discord or the repo or the internet for them)

personally for vehicle sim I would probably start with a raycast-based vehicle controller

rapier has one here
https://github.com/dimforge/rapier/blob/master/src/control/ray_cast_vehicle_controller.rs

there's also this
https://youtu.be/CdPYlj5uZeI

you can make one with raw physics simulation and contacts too, but it'll probably just be more painful and you'll have less control

[frail robin]

https://github.com/avianphysics/avian/issues/969

[prisma raft]

Basically I have ported it to the bevy and avian, but it breaks on high speeds. Thats why i decided to switch to ccd with real wheels

I think it can be solved with several raycasts, but it would require some sort of intermediate step just before the collision

[prisma raft]

Looks like its a bug https://github.com/avianphysics/avian/issues/705
Unfortunately its a show-stopper, with extremely high compilation times i have to switch back to c# 🙁

[sleek thicket]

did you FIX_INTERNAL_EDGES on that mesh?

[echo parcel]

Can't I have both Collider and Mesh3d in the same query?

[prisma raft]

It helped just a little bit, there are a lot of issues with instability. I've tried godot collision engine, and behaved much better, pretty stable on any meshes

[nimble fiber]

Yeah of course they're unrelated. Share your query?

[echo parcel]

pub fn dig(
    _: On<Fire<Dig>>,
    camera_query: Query<(&ChildOf, &GlobalTransform), (With<LocalEntity>, With<Camera3d>)>,
    spatial_query: SpatialQuery,
    mut chunk_state_monitor: ResMut<ChunkStateMonitor>,
    mut mesh_data_query: Query<(&mut MeshData, &mut Mesh3d, &mut Collider)>,
    mut meshes: ResMut<Assets<Mesh>>,
    mut gizmo_assets: ResMut<Assets<GizmoAsset>>,
    mut commands: Commands,
) -> Result {

[nimble fiber]

...Yeah nothing out of the ordinary. What happens? Does the query just not return anything?

[echo parcel]

I'm getting an error when I start the game. This started to show up after I've added Collider to the query

error[B0001]: Query<(Position, Rotation, Collider), ()> in system neklurgle::game::plugins::player::input::inputs::dig::dig accesses component(s) Collider in a way that conflicts with a previous system parameter. Consider using `Without<T>` to create disjoint Queries or merging conflicting Queries into a `ParamSet`. See: https://bevy.org/learn/errors/b0001

[nimble fiber]

Okay I was thinking it was going to be a conflict of some kind

[plain lion]

SpatialQuery conflicts:
https://docs.rs/avian3d/latest/src/avian3d/spatial_query/system_param.rs.html#60-64

[echo parcel]

What can I do about it?

[nimble fiber]

Essentially what the compiler error says

Your mesh_data_query queries Collider

So does SpatialQuery

So you either need to colocated both in a single ParamSet, use a QueryLens on SpatialQuery to extract the Collider query and remove it from mesh_data_query or... Find a way to use Without on mesh_data_query.

This is one of the parts of Bevy that is difficult and annoying.

Might be able to advise better if you share the whole system but no guarantees. Sometimes the right option here is to break things up into multiple systems

[echo parcel]

Here is the whole system

[echo parcel]

Ok, I went with the ParamSet

[somber portal]

I'm trying to implement a simple door opening animation from Blender. For some reason, the animation isn't moving the collider when playing, only the visible mesh. Why might this be the case?

[sweet sundial]

'regular' animations are gpu-side only

[somber portal]

Define "regular". As I have an old map with an animation that still seems to work without this problem. I've yet to find out why

[sweet sundial]

the type you'd load from a gltf or similar

what's the old animation? could be that a root bone will be read back or something, or maybe it had special handling

[glossy holly]

can you use position and rotation in non physics objects?

[somber portal]

Here's a screenshot of the working door in Blender:

The collider seems to move with the animation in this one

opposed to

This one, where the collider stays still:

I can try and provide short videos demostrating

[sweet sundial]

where's the collider in that second hierarchy?

[somber portal]

The second one i'm trying to just use the mesh of the door. The first one only has a seperate collider mesh because I was testing something a year or so ago.

[sweet sundial]

https://docs.rs/avian3d/0.6.1/src/avian3d/physics_transform/mod.rs.html#254-257 if they have https://docs.rs/avian3d/latest/avian3d/physics_transform/struct.ApplyPosToTransform.html

[glossy holly]

i see

[jaunty bolt]

I can't tell if someone else has already reported this as an issue, or if it might be my code, but I'm using apply_force_at_point with two points on the opposite side of the center of mass, as shown, at points that follow points on the body, dependent on rotation, and such that the resultant force should produce 0 torque, but which results in a torque that acts to align the body's rotation with the direction of velocity while moving with a high amount of velocity

something that seemed to mostly fix the issue was adding the body's velocity * dt to the point's translation, which leads me to believe that the transform is being updated somehow before the physics engine's center of mass is, despite me setting the body's transform anywhere

[orchid olive]

Are you doing your physics update in FixedUpdate, because that's where avian runs, and it can cause those kinds of problems when your running your physics logic on Update

[glossy holly]

is there a way to get the collision impulse magnitude before the collision happens inside of modify_contacts?

[vestal minnow]

Previous impulse yes, current impulse no. The solver is what computes and applies contact impulses, and it has to run after collision detection

However there's a normal_speed associated with each point

[glossy holly]

I'm trying to test a breakable object that doesn't collide if the object is fast enough

like angry birds ice

[vestal minnow]

That tells you how fast the shapes are approaching or separating at that point

which can be used to emulate "contact strength", ignoring mass

impulses can be a little fiddly since the time step affects them and I believe speculative collision can also lead to impulses being a bit dampened sometimes

[glossy holly]

so normal impulse would be the previous, and normal_speed is the current?

and if i add together all the points' normal_speeds * mass it should be some form of magnitude?

[glossy holly]

then probably divide by all the points?

idk

[vestal minnow]

normal_impulse is the latest computed impulse. If you were reading it after the physics step, it would be the "current" impulse. But you're reading it during collision detection, before the solver has run, so it's effectively the impulse from the previous timestep, not the impulse of the updated contact

Adding together normal_speeds is maybe a bit sketchy, you probably just want to pick the largest one

i.e. "how fast are these shapes approaching each other"

[glossy holly]

i see

[vestal minnow]

(or separating from each other if they were overlapping)

it depends on what you want as the heuristic for how "hard" the objects hit each other

[glossy holly]

i see

tyty

[echo parcel]

I need some knowledge. Is it the cube made of voronoi cells or is the cube "voronoied" at realtime?

[jaunty bolt]

that'd probably be it then

worked, thanks

the effect I described is still happening a little bit, but it's mild enough that I'm not that worried about it

[hoary turret]

"voronoied" in real time, also remeshed in real time to the voronoi cell shapes

[echo parcel]

Do you have any public repo where can I see how its done?

[hoary turret]

private for now, but I'm thinking about making it into a crate if I make it generic enough

right now it fits a very specific usecase where colliders and meshses are basically 1:1 and also they're simple low-poly shapes

which wouldn't work for most games really

[dull dew]

Is there a way to debug spatial_query.cast_shape with gizmos or something that will let me know where my shape is?
I'm casting a cylinder atm, i've tried drawing with gizmos.circle but i just can't get it to make sense and lign up

[echo parcel]

Yup, exactly how it works in my case

[echo parcel]

What movie is it from?

[vestal solstice]

labyrinth

[echo parcel]

1986?

[vestal solstice]

with david bowe

[unkempt jewel]

Hello 👋
I'm having some trouble trying to use point projection with a trimesh collider. Specifically the ColliderConstructor::TrimeshFromMesh .
If I use a cuboid or ConvexHullFromMesh the spatial query correctly determines is_inside but with a trimesh it seems to consider everything as being outside. Is this just a limitation due to trimeshes being an arbitrary shape compared to a convex hull that guarantees being a closed shape? or is there something fishy going on under the hood?

for context i'm just using a bevy cuboid as the mesh for the collider and want to get the distance from an arbitrary point to the closest surface on the collider

[vestal minnow]

Triangle meshes are hollow, they don't have any solid volume even if they are closed

Conceptually it's just a bunch of infinitesimally thin triangles that form a mesh

[unkempt jewel]

ahh ok, gotcha
any pointers on finding out if a point is "inside" a trimesh using the API? i didn't have good results with convex decomposition but if that's the right way then i'll rush headlong into that instead

[zenith river]

Are separate simulation contexts implemented yet?

[patent depot]

I remember seeing discussions about this.

[nimble fiber]

Oh man I remember this algorithm from like 10 years ago-

https://forums.autodesk.com/t5/3ds-max-programming-forum/how-to-check-if-a-point-is-inside-a-closed-mesh/m-p/13130768

The tl;Dr is you have to random raycast to infinity (or at least the AABB boundary of the mesh) and count the number of intersections and normals of those intersections. This is a raycast through all, not one with a limited number of hits.

This is only guaranteed to work for manifold meshes.

There's a much faster check if you have the SDF of the mesh but most people do not.

[nimble fiber]

If I recall Avian doesn't natively support them yet but you can use multiple SubApps/multiple worlds.

Or... Maybe there was some limitation there too. My memory's foggy.

[unkempt jewel]

This is great, thanks!
Incidentally I do have the SDF for some meshes as I wanted to use the point projection to cut through it. Figured projection would be easier than generating an sdf for an arbitrary mesh, but I'm gonna spelunk into that as well to compare both methods. Exciting!

[nimble fiber]

i mean, checking for whether or not an SDF contains a point is trivial

it's just sdf(my_point) < 0.

[unkempt jewel]

Yea. I hope it's similarly straight forward to generate an sdf

I can't see why it wouldn't be if I use both point projection and the raycast check you outlined

[nimble fiber]

Well- Usually meshes are generating from SDFs, not the other way around

Or you happen to have data which is mathematically equivalent to an SDF

that's why I said "most people don't have one," because unless you're creating SDFs for terrain or some kind of generative geometry, it's usually easier to just do the raycast check as opposed to try and back out an SDF from a mesh

[glossy holly]

I have enhanced determinism enabled, but collision events don't seem to maintain the same ordering

it's really weird, it alternates between entity 318 and entity 320 colliding first

seemingly every other run

[glossy holly]

adding/removing ColliderDisabled seems to alter the order of collision events, very strange :(

removing collision events and disabling the rigidbody instead fixes this, but very jank solution

[nimble fiber]

is there a meaningful difference between RigidBody::Static and RigidBody::Kinematic for Sensors?

[slim crypt]

Hello, is there a reason why CollisionLayers is marker as an Immutable component ? should we not edit it and re-insert it to override it ?

[nimble fiber]

the idea is that if you want to change it you insert a CollisionLayers component on an existing entity and it replaces the previous CollisionLayers.

[vestal minnow]

Mainly that the static rigid body is static (cannot have velocity etc.), while kinematic bodies can have velocity

[vestal minnow]

It's to minimize change detection overhead, for most apps I would expect modifying collision layers to be a very sparse and rare operation, and so having systems iterate over all colliders to see if their layers changed is unnecessary overhead

By reinserting and having observers react to it, the change is also applied "immediately" for spatial queries
etc., unlike when using standard systems with change detection

But we might revert this and make it mutable again if/when pcwalton's change indices land, since that should make the overhead negligible

[slim crypt]

Thanks ! I'm migrating an old game and this surprised me. I would argue that altering collision layers is more frequent than we expect, even though probably not on a large scale.

An other question, there used to be a ExternalImpulse component I could insert in a command to add impulse velocity to a rigidbody, are we expected to use the Forces query params ? Is there a way to use this in a Command?

[vestal minnow]

Yeah, Forces is the main force and impulse API now... in a Command (or instead of a command) I would maybe consider running a one-shot system via run_system, that way you can easily use system parameters and queries

in theory you can also cache a SystemState with the query in a resource or something, and in the command, access the system state in a resource_scope, but that's pretty inconvenient

[slim crypt]

Alright! I think I need to rework the associated logic anyway

[still hazel]

is there a way to check if rigidbody is on floor/wall/ceiling
or does it require manual mathematics

[nimble fiber]

nothing built-in as far as I'm aware, but the math isn't hard, lmk if you want a high-level description.

If you need this for a character controller, I'm pretty sure bevy_tnua has this built-in.

[small moss]

Would it be possible / how difficult would it be to use avian for collision detection, but use custom code for integration and collision response? Is there a sub-dep that I should be using directly?

[nimble fiber]

For #1, the general consensus is that you can use Parry (the collision detection library Avian uses under the hood) directly if you want, but it has no ECS bindings, so the API will feel very much so like you're calling out to a separate library, instead of like a bunch of components and systems that integrate with Bevy. You will also have to write glue code to, for example, synchronize entities/components with Parry's data structures (which is what Avian does under the hood). Jondolf's also working on an in-house collision detection library intended to replace Parry one day, but it's not done yet, FWIW.

For #2: Depends on what exactly you want to do.

Kinematic RigidBodies aren't the same as custom integration but if all you need is to directly control the position/velocity of specific bodies, and manually control how they are affected by other bodies, they're the simplest. The only tricky part is you don't really have fully direct control over how the Kinematic body affects other Dynamic bodies.

You can use CustomPositionIntegration and CustomVelocityIntegration markers to disable built-in integration for Avian on a per-entity basis. (I think) you'll want to add your own systems in the same place in the schedule schedule as the built-in integrate_positions and integrate_velocities systems which perform custom integration for those and only those bodies.

Alternatively you could take a look at the plugins added by PhysicsPlugins, and remove the IntegratorPlugin entirely if you want to do custom integration for every single body- You'd probably want to start by copying the code for IntegratorPlugin and modifying it. This is the nuclear option if your game's entire dynamics differ from the default lifelike ones Avian provides.

[heady oasis]

sooo.. I have two entities and I wanna get entity 1's position when they collide

[nimble fiber]

if you just want which two entities are colliding and only need the like, origin position: https://docs.rs/avian2d/latest/avian2d/collision/collider/struct.CollidingEntities.html or https://docs.rs/avian2d/latest/avian2d/collision/index.html#collision-events depending on whether you want a system or observer.

if you want the actual point at which the two entities are touching: https://docs.rs/avian2d/latest/avian2d/collision/contact_types/struct.Collisions.html

[small moss]

The plan is to use my existing position/velocity system for all objects, I just need to know when they collide.

[heady oasis]

all I want is just to place a .observe on entity 1
check when it collides with entity 2
do stuff with where entity 1 ended up

ooor loop through all collisions that happened this tick
check if collider1 or 2 is entity 1 and vise versa
"collision happened"
but that feels like a bad aproach tbh

first thought

// .observe
query : Query<&entity1> // error[E0277] the trait `IntoObserverSystem` is not implemented
// then do 
if query.contains(collider1){
  // get position out of query this way
}

[nimble fiber]

you could go either with parry or avian for this. Parry has no built-in broad phase, so if that matters to you you'd need to reimplement it which isn't trivial, but otherwise it has what you need. With Avian, if you're working with position/velocity and want objects to depenetrate, i.e. if two objects touch they don't enter one another's collider, you could use Kinematic bodies which have velocity/position set manually. You could also use sensors if you want no dynamics whatsoever and want to granularly control everything yourself.

[nimble fiber]

fn my_system(query: Query<Entity, &MyMarkerComponent, &CollidingEntities>) {
  for (entity, my_marker, colliding_entities) in query {
    for other in colliding_entities {
      // do stuff
    }
  }
}

[small moss]

Broad-phase was one of the problems I was hoping to farm out.

[nimble fiber]

I'd stick with Avian then. If you are using none of the dynamics (i.e. no RigidBody::Dynamic), then the performance cost for the "extra parts" of Avian will be nil. Again, you can always put together your own plugin that excludes systems related to physics engine dynamics. The only concern would be extra binary size/compile time but unless your deployment target is a gameboy it won't matter.

[still hazel]

do static rigidbody have forces

are static rigid bodies included when querying forces

[still hazel]

if not

oh wait i could just
forces_query: Query<Forces>,
rigid_body_query: Query<&RigidBody>,
and does forces.apply_force(force) work on kinematic rigidbody

[bleak mirage]

@vestal minnow ran into a weird obscure bug with avian2d.

I have a scene full of static rigidbodies, and their global translation is updated almost every frame (seemingly by avian, since if I remove the rigidbodies altogether it stops being updated), even if no entity in the scene is moving. It's a huge problem because their z is updated as well, and it seems to jump around a bit, e.g. entities have it set to 50.0 when they spawn, but some frames it changes to like, 50.00001 and then back the next frame.

I'm sorting the objects by their z and this is causing massive visual glitches :)

I feel like, at the very least, avian2d shouldn't touch or do any computations with the z value

hm apparently interpolation might be the culprit; @quick sluice can provide more details, it's their project that I encountered this on

[quick sluice]

So the visual glitch does not exist when the PhysicsInterpolationPlugin is set to default. It exists when the translation is extra - or interpolated.

[quick sluice]

I am using avian2d.

[stiff cypress]

Hi, is there a way to change entity from static to dynamic rigidbody? Doing the other way around works perfectly with this code

commands
    .entity(root)
    .remove::<RigidBody>()
    .insert(RigidBody::Static);

and trying to do the opposite like so makes entity ignore all the colliders and infinitely fall

commands
    .entity(root)
    .remove::<RigidBody>()
    .insert(RigidBody::Dynamic);

[sweet sundial]

if you remove the .remove::<_>() line does it work

[stiff cypress]

hmm, it works this way, but that is not my actual issue. What i try to do is separate one dynamic body (root) without collider but with children that have colliders into 2 dynamic bodies, i initially used cloning but removed it to find the source of the issue. here group are those entities with colliders but without rigidbodies

                let new_root = commands
                    .spawn_empty()
                    .insert((
                        RigidBody::Dynamic,

                    ))
                    .id();
                commands.entity(root).detach_children(&group);
                commands.entity(new_root).add_children(&group);

and it still have the same issue as if i was removing and inserting

[sweet sundial]

what if it's let new_root = commands.spawn(RigidBody::Dynamic).id();?

[stiff cypress]

the same

it might be related to colliders Bounding Volume Hierarchy (BVH) or Islands

[stiff cypress]

is there a way to temporary disable those to see if it fixes the issue?

also sometimes both old and new rigidbodies break and then one fall down and the 2nd one flies up

[peak timber]

There can only be one component of a given type per entity so removing is redundant

[stiff cypress]

Ok, so i was able to fix the issue via setting dynamic after reparenting; via setting dynamic in different system like so

                let new_root = commands
                    .entity(root)
                    .clone_and_spawn_with_opt_in(|builder| {
                        builder.allow::<(Name, Transform, Visibility, RigidBody)>();
                    })
                    .insert((TransitioningToDynamic {},))
                    .id();
                commands.entity(root).detach_children(&group);
                commands.entity(new_root).add_children(&group);

#[derive(Component)]
pub struct TransitioningToDynamic {}

pub fn transition_to_dynamic(
    mut commands: Commands,
    ent: Query<Entity, With<TransitioningToDynamic>>,
) {
    for entity in ent {
        commands
            .entity(entity)
            .insert(RigidBody::Dynamic)
            .remove::<TransitioningToDynamic>();
    }
}

I am also not able to replicate with minimal setup

#[derive(Component)]
struct Root {}
#[derive(Component)]
struct Red {}

fn separate(
    mut commands: Commands,
    keys: Res<ButtonInput<KeyCode>>,
    root: Single<Entity, With<Root>>,
    red: Single<Entity, With<Red>>,
) {
    if keys.just_pressed(KeyCode::Space) {
        let root = root.into_inner();
        let new_root = commands.spawn((Root {}, RigidBody::Dynamic)).id();
        let red = red.into_inner();
        commands.entity(root).detach_children(&[red]);
        commands.entity(new_root).add_children(&[red]);
    }
}

fn setup(
    mut commands: Commands,
    mut meshes: ResMut<Assets<Mesh>>,
    mut materials: ResMut<Assets<StandardMaterial>>,
) {
    commands.spawn((
        Root {},
        Transform::from_xyz(0.0, 4.0, 0.0).with_rotation(Quat::from_rotation_x(1.)),
        RigidBody::Dynamic,
        children![
            (
                Mesh3d(meshes.add(Cuboid::new(1.0, 1.0, 1.0))),
                MeshMaterial3d(materials.add(Color::srgb_u8(124, 144, 255))),
                Transform::from_xyz(-1., 0.0, 0.0),
                Collider::cuboid(1.0, 1.0, 1.0)
            ),
            (
                Red {},
                Mesh3d(meshes.add(Cuboid::new(1.0, 1.0, 1.0))),
                MeshMaterial3d(materials.add(Color::srgb_u8(255, 144, 144))),
                Transform::from_xyz(1., 0.0, 0.0),
                Collider::cuboid(1.0, 1.0, 1.0)
            )
        ],
    ));
}

[stiff cypress]

oh, actually when i look at it, it inserts dynamic for dynamic too and that actually seems to fix those broken bodies

[prisma raft]

It seems to me that it its easier to make analytic collision solver for my special case, rather than try to fix mesh collisions. The reasons are:

• sphere colliders are literally stopped by ghost collisions even in a situations they are not supposed to collide, and this is 100% reproducible;
• fixing meshes does not helps;
• raycasting does works, but with caveats;
• for wheeled car, raycastsed "wheels" can stuck in the geometry on high speed, and without any tweaks are unable to recover;
• prismatic joints are not solving this problem at all, creating ghost "pull" forces, and pushing car with mass into the ground;
• the only way to solve this, is to create "kinematic" objects for the "wheel" joints, but this looks awful;

Maybe I'm doing something wrong, but I've tried different methods of handling and without any noticeable success. The last idea was on the ray cast enable collider, so prismatic joint would work correctly, but that would make system even harder to maintain

Basically i need hermite splines, they can be subdivided to find right "t" position, then i need reprojection into the extrusion slice, and I'll get correct solid, rather than hard-to-handle mesh

[sleek thicket]

wdym stuck in the geometry

[prisma raft]

Imagine that wheel travels from "not inside track" into "inside track", raycasted "wheel" must apply correct impulse to move car chassis such that wheel will be lifted into the correct position above the surface. There can be a special case when raycast from the wheel position will be inside "ground" too.

Mesh is not solid, so I either have to create more raycasts from above, or do something else

let me make a screenshot

Basically it looks like this, wheel raycasts cant push car from the ground fast enough, so they are clipping

[sleek thicket]

you mean the wheel model clipping?

[prisma raft]

yes, and that should not happen. there are cases when on the high speed whole wheels with raycast points go under the track, and there are no ways them to recover, because they are under the mesh, and cant register any collisions

[sleek thicket]

especially this

[prisma raft]

Thats exactly why I'm complaining about it.
I cant use swept spherical colliders for the wheels, if they worked properly, that would solve all problems

[sleek thicket]

do you need the wheels to be specifically outside of the car?

[prisma raft]

red box is a cuboid collider with swept ccd. and as you can see, even this dont solve issues with incorrect collision registration

[sleek thicket]

yeah, but you can at least reduce the issue by making the rays start from inside that collider

[prisma raft]

it looks like I'm pushing current physics system waaay beyond its capabilities. analytical collisions would solve this task perfectly well, but I dont know where to start

[sweet sundial]

try using shapecasts

[prisma raft]

i do spherecasts right now

[fathom cedar]

for anyone doing car physics, this is interesting https://youtube.com/shorts/lsaTt030jVY?si=1AR8mFUwyZfX4m4c

[sleek thicket]

shapecast with extra steps

[dull dew]

yeah seems like more work both on compute and LoC
I refer people to this video if they want simple car physics
https://youtu.be/SKXqWcaoTGE?t=260

[frank sparrow]

hey everyone, wanted to ask if there is any way for a CollisionStart event to have a non-finite vec3 in avian3d in collisions.manifolds.first().points.first(). Not sure if it makes any difference but this would be for 2 kinematic objects with one having a much higher mass than the other

[sleek thicket]

new video from 1 of developers of very very valet

@dull dew @prisma raft @fathom cedar

[dreamy lion]

Are there any areas where I can reduce the performance overhead of the physics engine itself? I am trying to run around 500 bevy worlds with each 200 objects (all in one process) and it seems like the cpu spikes mainly because of the physics overhead itself since 50 worlds with 2000 entities work fine. I have the exactly same result with rapier

[prisma raft]

parry supports sphere raycast out of the box, there are no problems with raycasting itself, but there are no reliable ways to do raycasts on high speed, such that bumps would be correctly simulated

[prisma raft]

already did that, im stuck with physics stability and collision detections for (at least) shperical colliders, that are not properly working

as I already said:

• I cant use spherical colliders because of phantom collisions/forces;
• I can simulate spring forces, BUT, there are two both non working ways
  • prismatic joints are not supposed to work with the simulated raycasted "wheels"
  • custom simulated joints suffer from ground clipping

it might work with simple scenes, but dont work at all with fast paced rally / racing game

Im using procedurally generated roads from the hermite splines, that are baked into the meshes at runtime. I cant do it offline (in any editor like blender or houdini), because level editor is supposed to work inside app

[sleek thicket]

raycasted wheel is already a custom primatic joint, why would you combine them? or am i misunderstanding something?

[prisma raft]

Thats the essential question. The main problem is that car chassis must be pushed out of any collision surface without applying impulse and momentum. Its constraint solving, that is available only with joints. I cant just apply any (basically very large) impulse to move chassis out of the ground for exactly one frame and immidiately cancel it out. This also true for momentum

It basically a hard stopping constraint, that cant be simulated using phantom forces

[sleek thicket]

why do you want to push chassis for 1 frame, do you need to make your car without any suspension?

[prisma raft]

I need suspension to work like:

• spring to absorb any impacts;
• damper, to avoid any oscillation;
• be preloaded, such that it will already have some force applied;
• also have HARD STOP (bumper), such that it cant be pushed any furhter

in case when suspension reached in min limit, it cant be pushed any further, and this is a hard constraint

only in this case car would not dive under the collision sufrace

[sleek thicket]

...but all of that should already be handled by raycasted wheel

[prisma raft]

no

it would not, because raycasted wheel does not have constraint on its min lenght

[sleek thicket]

but it does?

[prisma raft]

how?

[sleek thicket]

how does a spring not have a min length constraint?

or are we thinking of different things?

[prisma raft]

how you can ensure and apply force / impulse to correct situation when wheel is under the ground on the next frame?

[sleek thicket]

why would it be underground in the first place?

[prisma raft]

because car is moving, and environment does have bumps and ramps. if car is moving fast, then without swept ccd you cant detect situation, when wheel is under the surface

thats why swept ccd exisits in the first place

this is kinematic picture. if wheel is under the ground, you cant just move car upwards without applying any force or impulse. its not how physics works

to avoid situations like this you have to apply hard constraint, such that movement beyond this limit is prohibited

[sleek thicket]

when wheel hits bumps or ramps, it moves up, to whatever point shapecast hits, or do you mean the flat bumps where wheels would just look weird if they teleported?

[prisma raft]

how can you teleport wheel? i mean in physical sence, which impulse should be applied?

you can basically make wheel a kinematic body, such that it will have constraints, but thats extremely hacky and buggy

whats the point of physical engine in this situation?

[sleek thicket]

wait, what point do you cast your shape from?

[prisma raft]

it does not matters, you can cast just from chassis

[sleek thicket]

so are you trying to solve a visual bug, not physics?

[prisma raft]

its a physical bug in the first place

the only question is, how do you correct wheel and chassis position:

if due to high speed and without swept ccd wheel goes under the ground reaching below allowed min suspension lenght

you cant just apply impulses and forces, because it would work like a phantom force, and look ugly

[sleek thicket]

so we've been talking about different things after all 🫠 you're thinking about a collider, while i'm thinking about shapecast

either way if you'd run into a bump that's so high your wheels would have to be underground, the car should flip over instead

[prisma raft]

shapecast should behave exactly as collider in order to be consistent

there must be no frames then wheel visual is under the ground because shapecaster and suspension didn't had enough time to push car away from the ground

i mean, im not trying to write toy example. i want to make it look and work realistically

[sleek thicket]

what kind of car though? in realistic scenarios you'd either crash the car itself, or the bump would just make you fly up

[prisma raft]

here is how it was fixed for kinematic bodies, its move and slide:
https://joonaa.dev/blog/12/avian-0-6
its basically move-and-slide case

[prisma raft]

If you really interested in digging into this case and helping me out, please write your own car model, and make it work. Otherwise i dont see any point in explaining it any further

for the reference, speed limits are like those: https://www.youtube.com/watch?v=zTYJJad6aJQ

[sleek thicket]

i see, the ray/shapecast wheels weren't made for this case at all, but trackmania is a good example of how sticking to 1 genre for 20 years allows you to refine the mechanics 😄

[fathom cedar]

I watched this yesterday! It's a good video breakdown of arcade style car physics. Did not realise he worked on very very valet.

[prisma raft]

English isn't my first language, so I want to understand exactly which parts of my explanations and clarifications might have been unclear.
I've tried to describe a specific issue with the engine that, so far, can't be resolved without writing my own physics handler. It is possible to extend the engine strictly for my specific needs, but I don’t yet know where exactly to start

[sleek thicket]

they weren't just unclear, they were misleading 😅
i'm not sure how you'd extend it without fully understanding it first, so i think that'd be the start?

[prisma raft]

I've just re-read it again, and cant find any obvious flaw aside that actual case (very high car speed) is not mentioned

[sleek thicket]

you were re-reading while understanding the context, and what your special case is

[prisma raft]

Can I take this to mean that the question was phrased in too technical a manner, without explicitly explaining why other approaches wouldn’t work, given the task’s complexity—which is unusual for hobby projects?

[sleek thicket]

no, it means only what it said, your imagination filled the gaps while reading my message, just like your imagination filled the gaps while asking the questions 🫠

[prisma raft]

there are several quite interesting points on how to handle ground penetration, overall its great video, thanks!

[torpid hinge]

I'm new to Avian- are there tools that would help me visualize the trajectory of a physical object? Think parabola of something shot out of a cannon while affected by gravity.

It's for debugging so happy to use bevy gizmos for showing the trajectory, but I'm wondering if Avian has anything to calculate paths like this.

I think perhaps this is equal to simulating the rigid body's position for the next N ticks?

[nimble fiber]

I don't know that Avian really has built-in utilities to make this any easier- What's your usecase?

[torpid hinge]

I'm testing Avian and I'm going to shoot some heavy projectiles (i.e. cannonballs) with some rng, and I'd like to be able to debug visualize their trajectories.
Could also be nice in a type of game where you plan to "lob" some object in an arc, and you want to let the player know the trajectory (pool/snooker, golf, ..)

[nimble fiber]

So, I think most games do this by just using a shapecast (or even a raycast) that emulates the properties and simulation of the rigidbody

If it's parabolic all you have to account for is gravity, possibly with linear drag. You'd probably want to lock the rotation of the rigidbody entirely.

I'm not really aware of any physics engines that let you like, step a rigidbody into the future n times.

If you REALLY want a high-fidelity prediction I wonder if you could do a partial clone of the world and then run the physics schedule n times. But that's a lot more involved than a shapecast.

[bleak mirage]

there are a few crates that do this in one form or another, mainly network crates which do it for rollback (part of rollback is stepping N ticks into the future, including physics)

I’ve also seen this standalone crate that seems to do exactly this: https://discord.com/channels/691052431525675048/1488695835486326866

but yeah an actual separate physics simulation to show the trajectory is not usually worth it, unless it’s a very core mechanic of the gameplay and there are too many factors into play to just determine it directly

[torpid hinge]

It's not a core mechanic, I just hoped there were some helpers for this in Avian to avoid papercuts like e.g. a future refactor in 6 months which changes linear drag somewhere else, gravity, or some unknown unknowns..

[nimble fiber]

Yeah afraid the answer is no, and I think that would be the case for most physics engines frankly. I think it would honestly be a pretty cool feature though, like a predict_step of some kind where you can plug in various properties to get a prediction assuming the rest of the world is static.

[torpid hinge]

Yep exactly, "where is this in N ticks assuming no other changes" would be very nice and help avoid creating crude emulations

[glossy holly]

yeah but i find that for a lot of games it's not just running physics but also other collision systems and such

which would be impossible to do predictively without proper simulation?

like a pinball bumper

[nimble fiber]

Agreed which is why I'm guessing most physics engines don't have it.

Though for pinball you could use the "simulate the entire world forward by n frames" approach that the various networking libraries use.

[glossy holly]

yeah, at that point it would need to be entire world simulation

which wouldn't be relevant for physics

but it could be interesting maybe to atleast have a "project body x secs later"?

hmm, in the cannonball example it wouldn't help much i feel

[nimble fiber]

yeah I think you really have 2 practical choices in the vast majority of games:

• shapecast n times, syncing up the length of the shapecast with the length of the delta, and reusing as much "game logic" as possible to make it mostly match how the rigidbody will behave
• clone-the-world and simulate into the future

I don't think a middle path really exists. I think the most useful thing would be to expose a RigidBody::integrate to make it easier to clone the integration of Lin/Angvel.

[glossy holly]

i think in cannonball case you don't need to simulate game logic too much

but you do need some sort of shapecast arc?

could also be cool for pool indicators

or maybe jumpking style rigidbody games?

pogostuck

[glossy holly]

trying to make radial gravity causes an infinite speed loop

this is following the example here

i believe it's because the acceleration is applied at the same direction throughout the timestep, and so when the ball moves + acceleration staying the same causes a push

possibly like this

reducing subset count seems to reduce but not mitigate the effect

i think true radial gravity would need a apply_linear_acceleration_towards(Vec2)

changing this to directly modify the linear velocity seems to work, as it's not applied continiously

[nimble fiber]

what is direction.truncate()?

I see trunc() in the method docs but not truncate()

which appears to integer-ize each component by rounding down

[glossy holly]

yeah not sure why that is necessary?

[nimble fiber]

maybe remove it?

[glossy holly]

i don't have it

[nimble fiber]

oh that's the example my bad

[glossy holly]

yup

that's in the docs

[nimble fiber]

I thought that was your code lol

[glossy holly]

yeah

[nimble fiber]

if you're happy with the linear velocity approach go nuts, happy to read over your code if you want help with using Forces

[glossy holly]

the forces code is quite similar

just applying acceleration towards the center

im quite confident it's not possible to do with forces without modifying avian

[glossy holly]

im pretty sure this is the situation

[nimble fiber]

I wonder if this is because of substeps

[glossy holly]

reducing substeps seems to reduce the effect

[sweet sundial]

you could put the gravity inside the integration set and use the velocity as a second input

[nimble fiber]

i almost wonder if this is something Avian should publicly expose as part of the API as this feels like a fairly common usecase that should be straightforward to get correct

Like a way to override or add to rigidbody integration like you can in Godot

[glossy holly]

what does this mean?

[sweet sundial]

order the gravity system between something in the avian 'internals'
and use a gravity approximation that understands current velocity

[nimble fiber]

@glossy holly https://docs.rs/avian2d/latest/avian2d/dynamics/solver/schedule/enum.SolverSystems.html#variant.PreSubstep

substepping isn't like an internal engine black box, it's a system set you can put your own systems in

you have to be careful to schedule them at the right time in the substepping loop

but it would let you access the actual position/velocity of the rigidbody during the substeps and update the direction of radial gravity to reflect the actual position of the body during substeps

yeah I feel like Avian should be exposing a public CustomIntegration schedule which you can assign systems to which run either before or after built-in engine integration, and maybe a DisableIntegration marker component to disable built-in integration or something like that

I think SubstepSchedule in particular is the one you'd want to run systems like this in

[glossy holly]

yeah im messing around with that but its breaking some of my other effects

[vestal minnow]

we have a CustomPositionIntegration component, which is useful for e.g. custom movement logic using MoveAndSlide, but we don't currently have a CustomVelocityIntegration component

[glossy holly]

trying to run it before ApplyConstantForces currently applies a crash

[vestal minnow]

actually no

[glossy holly]

hi john physics

[vestal minnow]

we do have CustomVelocityIntegration :P

CustomVelocityIntegration and CustomPositionIntegration

[glossy holly]

hmm

that seems much more involved than the linear velocity solution

[vestal minnow]

yeah

[glossy holly]

would it ever make sense to implement an "apply_force_towards" another body?

in engine

oh wait

i bet i could use a fixed joint with point compliance for gravity too if I wanted

lmao

[vestal minnow]

that seems quite game-specific

[glossy holly]

yeah im fine with the velocity solve

the substepschedule solution seems to have problems

[vestal minnow]

We should definitely add some examples for implementing radial gravity and gravity fields and whatnot properly

and possibly some first-party functionality

similar to Godot's area influence

[nimble fiber]

Oh I totally missed those components

[vestal minnow]

yeah scheduling in the internal schedules can be kinda tricky and also can change a lot between releases

[glossy holly]

if that wouldn't be proper, what would be?

[glossy holly]

referencing here

[glossy holly]

maybe having a PreSubstep/PostSubstep schedules could allow this functionality, and be more universally useful?

[nimble fiber]

they already exist and aren't in the substep loop so they'd need a different name...

It's like "pre loop" and "post loop"

[glossy holly]

oh true forgot about those

yeah

I think PreSubstep is a better name for inloop tbf, feels closer to PreUpdate

[vestal minnow]

Well it depends on the desired accuracy. I would start by just modifying LinearVelocity in FixedUpdate, but if that's not accurate enough (i.e. objects diverge from the orbit under radial gravity) then it could be integrated into the actual simulation loop, similar to what you're trying to do

[glossy holly]

linearvelocity is good enough for my usecase, but i think it is technically less accurate than applying a continuous acc or force

[vestal minnow]

it is less accurate but whether that matters depends on the project

[glossy holly]

yupyup

maybe this behavior can be documented?

[vestal minnow]

in most games you probably wouldn't notice

perhaps yeah

[glossy holly]

yeah but my original radial implementation following the example was quite noticable lol

[glossy holly]

troll physics

[vestal minnow]

yup that example isn't great hehe, like you speculated I think it's because the acceleration direction is only accurate for the first substep, since the direction remains constant across the substeps when it should really change as the shape orbits

and so the acceleration has a tangential component that speeds you up

[glossy holly]

even reducing it to a single substep still resulted in some acceleration, although less

[sweet sundial]

same error will still be there if you use the same naiive approximation yeah

iirc there's a substantially better approximation made by just swapping the order of operations somehow

[vestal minnow]

that's explicit vs. semi-implicit Euler (aka symplectic Euler)

semi-implicit is the stable one that most game physics engines use, Avian included

"stable" as in it preserves energy over long periods without artificially adding any energy

(but may dampen over time)

[nimble fiber]

Would you accept a PR that adds a warning describing the details of this to the radial example and Forces?

IDK how exactly to approach adding custom integration to the substep but some extra docs might be helpful for the next time someone encounters tyjks.

[vestal minnow]

I would maybe replace the radial gravity example for apply_linear_acceleration with something else altogether (maybe just a custom gravity field that points in some specific direction), since it can't really be used for good radial gravity as noticed by nukeexplosions

[nimble fiber]

Ahh maybe with a note about non-constant force fields?

[vestal minnow]

Yup it'd be good to have a warning about how the acceleration direction and magnitude remain constant across the whole physics timestep, which can result in inaccurate results for e.g. non-uniform gravity fields

[nimble fiber]

I could also try messing around with adding a dedicated SystemSet for running systems at specific points in the substep, but I don't wanna overcommit lol

[glossy holly]

I think it would also be good to document the situation where you would modify linear_velocity rather than apply_linear_velocity?

[nimble fiber]

Yeah for a more-accurate-but-not-perfect result that's easy to implement

[vestal minnow]

you mean apply_linear_acceleration?

[glossy holly]

yeah

hmm not sure how useful modifying linear_velocity for acceleration would be outside of this usecase

but having to figure that out was still a little obtuse i guess

[wet niche]

@vestal minnow I have noticed an issue where random meshes in my scene become visible. I tracked it down to avian3d::debug_render::change_mesh_visibility. store.is_changed fires a lot, even when no debug rendering is enabled (a UI system of mine is doing this). Should this function set visibility to Visibility::Inherited instead of Visibility::Visible, perhaps?

(Making that workaround works for me, btw. I can make a PR if needed) (I did it: https://github.com/avianphysics/avian/pull/981)

[nimble fiber]

If I modify or insert a Collider, I assume that I have to wait until the next frame for the Collisions SystemParam to be updated in the event that modifications to that Collider result in it either beginning to, or ceasing collision with another?

[echo parcel]

Is there any kind of 'Collider Instancing'? If I spawn e.g. 1000 of the same colliders will it make a difference over spawning 1000 of different colliders(let assume that the total amount of vertices, indices etc. remains the same)

[wet niche]

I don't think it's automatic but a lot of duplicates will make a difference. You can pool them manually by creating Colliders through Shapes that manually track a SharedShape. Avian uses Parry for shapes atm, so see https://docs.rs/parry3d/latest/parry3d/shape/struct.SharedShape.html which answers the questions in more detail.

[nimble fiber]

okay, more interesting question-

I know it's possible to run the entire PhysicsSchedule manually with world.run_schedule(PhysicsSchedule);

Is it possible to just run BroadPhase and NarrowPhase in order to update the contact graph?

I have a system that essentially wants to do all of the following in one frame:

1. Perform spatial queries (raycasts), and then modify/add colliders depending on the results of those queries
2. Recalculate the contact graph
3. Do 1 again, until all modifications/additions are exhausted and the contact graph is stable and needs no further modifications (or an infinite loop is detected, in which case there is a special handler)

I'm gonna start with just running the PhysicsSchedule as many times as I need to, but I know that I don't need to run Solver, Sleeping, and SpatialQuery. All of the Colliders in question are Sensors, so dynamics will never be affected by this particular system (though it can indirectly affect dynamics ex. by applying forces to bodies, that can wait until the next frame).

(to give a more tangible mental image- I'm trying to resolve Ghostbusters-style laser beam intersections in 2d space. If there's a bunch of laser emitters, mirrors, splitters, etc. I need to calculate what the network of intersections actually looks like)

[sweet sundial]

you could add a run condition to the unneeded systems;' sets

[nimble fiber]

I think I would need to make a copy of PhysicsPlugin for that right?

'cuz I'd need to alter the initial registration of those systems

[sweet sundial]

just app.configure_sets(PhysicsSchedule, PhysicsStepSystems::Solver.run_if(|r:Res<_>| r.should_solve()))

[nimble fiber]

okay yeah and I essentially just set that resource's value at the beginning of my exclusive system and toggle it back off at the end

okay that's pretty straightforward...

[glossy holly]

following this approach, would you apply buoyancy unevenly by maybe applying the force at the average of the sampled points?

[sweet sundial]

buoyancy forces can be applied at the center of displaced mass, which is pretty close

[glossy holly]

yup makes sense

[glossy holly]

water physics in avian2d

[glossy holly]

with visualization

maybe i should pr an example

[nimble fiber]

this is really cool

reminds me of a gamemode I used to play... I think it was in Team Fortress 2? Where the water would rise and the goal was to be the last one standing

may haveb een garrysmod

[glossy holly]

tyty

[sweet sundial]

might need more drag

[glossy holly]

yeah I added configurable angular and linear damping that scales depending on % of body in the water

[sweet sundial]

not going to have anything going quickly sideways on the edge?

[glossy holly]

I guess I could implement better drag but idk it's not super significant to my game

[vale dragon]

hey folks, wrote an Avian plugin for a Flight Dynamic Model: #showcase message

[vale dragon]

(I hope the crate name is ok, avian_fdm, it's not published yet, I'm happy changing the name if needed)

[velvet patio]

Anyone currently working on/maintaining an updated 0.19-dev branch? The one in PR#965 broke w/ the Bevy PR#24164 yesterday

[coral kite]

Is there any recommendations for how to do ragdolling with .glb meshes w/ skeletons in Avian?

Would I have to just spawn my .glb -> find the entity in the world -> put the colliders on the bones?

[coral umbra]

i'd imagine it'd be easier to use Skein and attach the colliders there

[round lion]

I am very new to Bevy/Avian so I apologise if this is obvious, but is there a convenient way to "subtract" a shape from a collider (eg. take a cuboid collider and cut the bottom of a sphere out of it to create a "dish" shaped collider)?

[glossy holly]

nope

[round lion]

Alright, thanks

[glossy holly]

i have a hacked together solution of something similar by using collision filtering

but it creates somewhat janky behavior

[round lion]

It was just for a tiny procrastination toy anyway, more reason to just get on with my actual project 😄

[nimble fiber]

i remember someone asking me something similar to this in the Godot server regarding a banana-shaped collider, I don't think any physics engines support this, but it does make me wonder what an SDF-based collision engine would look like and whether it would be really cool and useful or horrifically inefficient.

Maybe a project for the far future lol

[glossy holly]

ive seen a video about this actually

i think most people use sdf + marching cubes/squares approach

but sdf collisions are also possible

ive looked up on this before though

and subtract functions break the "validity" of an sdf

which doesn't work for pure sdf based physics

https://www.youtube.com/watch?v=il-TXbn5iMA

here's someone that did marching cubes approach

[nimble fiber]

so tl;dr is it's way more complicated than just using raw SDFs

[glossy holly]

https://dlpf.github.io/sdf-collision.github.io/

here's the research i saw on raw sdf collisions

from what i understand though, they sample points on the sdf and gradient descent to find collision points

but they preallocate those points

which also doesn't work w dynamically generated sdfs

tldr pure sdf physics works but doesn't work w dynamic subtracts and such

[sweet sundial]

someone had sdf colliders specifically in avian

[cinder summit]

Hi, I am someone 👋

My impl doesn't work on just any arbitrary combination of shapes. But if the goal is jist subtracting on some shapes rarely, turning it into a mesh is probably easier. Since then the rest of collision still work fine as-is

[glossy holly]

hihi can I see it

[cinder summit]

https://github.com/nisevoid/sdf_peck

[nimble fiber]

The more I do this the more I feel like a really good meshing algorithm ends up being the solution to a ton of problems in the videogame world.

Kinda like how a single excellent general-purpose database is often a better solution to most problems in the webdev world than a bunch of specialized options.

[plucky elbow]

@vestal minnow ya ever gonna make use of your reserve? https://crates.io/crates/avian

feels like should have the various avian's under it as features?

[velvet patio]

@untold peak are you planning to update your bsn branch to work on latest bevy soon? otherwise I might fork yours and fix the issues introduced last week so i can move to latest more easily

[untold peak]

These are my changes from a branch off bevy main a few weeks ago

I’ve been waiting for the dust to settle on bsn with 0.19 release (not a RC) before doing a full migration in jackdaw

[velvet patio]

All good, I just updated off yours, going to put a PR up for parry to update glam so we can make it more simple

[velvet patio]

I put a PR up for parry to update to newer glamx and have an updated avian branch (starting with cherry picking your fixes) using that — can put up a PR once I do a little more testing

[proven prism]

When is Avian going to be compatible with Bevy 0.19-rc1?

[vestal minnow]

Hopefully at some point yes, it's a bit awkward atm since the 2d and 3d features are mutually exclusive, so in that sense it's nicer to have them as separate crates

But if/when we refactor things such that 2D and 3D are more cleanly split and can be enabled simultaneously, it'd be neat to use avian

[vestal minnow]

I believe @velvet patio was working on a branch for it (see the message above yours)

I've been sick this past week and also have other work to do, so I haven't had much time or energy for it. But I'm starting to feel better now, so hopefully we can get an RC release done in the next few days

[velvet patio]

I have one working w/ latest git main but needs some patching in your workspace.to use the parry branch I have on glamx 0.2…

https://github.com/stuartparmenter/avian/tree/bevy-0.19

https://github.com/stuartparmenter/parry/tree/glamx-0.2

[proven prism]

Your health is what matters most — look after yourself first. Wishing you a speedy recovery.

[velvet patio]

I did put up https://github.com/dimforge/parry/pull/418 but it is just sitting there. Feel free to take my changes or I can put a PR up for avian if easier or helpful

[sour goblet]

Hey, Can I ask a question here? I have made my question on #1507360724102676671 message but not sure if someone who can maybe answer will notice over there. Thanks

[civic atlas]

Hope you get well soon. Remember that you are allowed to rest after getting sick 👍

[echo parcel]

I would like to save forces to the file and apply them to the player again when they load. How should I do it? I mean, just in avian, I have whole saving mechanism in place. What data should I save, and how should I reapply it again

[dreamy viper]

Hi, could I get a reminder on which of Position or Transform should be used? Should Position considered internal to avian and the user should always be modifying transform?

[cinder summit]

Correct. Either you use Transform, or you disable the Transform -> Position propagation

[bleak mirage]

funny seeing you ask this here since I look this up for Lightyear constantly and I’m pretty sure it’s the opposite haha

(since like, Transform is used internally for interpolation so you’re generally expected to change Position, or just use forces, might be remembering it wrong though)

[wet niche]

FWIW, there is a resource PhysicsTransformConfig where you can configure this logic. The default is to propagate Transform edits to Position and Position edits to Transform. along with a few other knobs.

[sweet sundial]

imo use Position inside your physics schedule and Transform in Update

[dreamy viper]

That depends on the avian mode used in lightyear. With mode=Transform, you are supposed to only interact with Transform. But I wanted to know what the recommendation is on the avian side

[vestal minnow]

In general, Transform in user code, and Position inside the PhysicsSchedule

[prisma bay]

hi

so I got hit a perf cliff with ColliderTreeOptimization defaults that I feel is worth flagging. App streams in a few hundred Static trimesh terrain colliders from a worldgen pipeline. Worldgen runs heavy stuff on AsyncComputeTaskPool (density compute, meshing, BVH builds).

Chrome trace at defaults:

PhysicsSchedule    total 36s (72% of update)
block_on_optimize…  mean 36ms / tick, max 730ms

I have 20Hz physics, small tree, low moved_ratio. The optimization work itself doesnt seem expensive, the cost is the wait.
optimize_trees spawns onto AsyncComputeTaskPool (default use_async_tasks = true on native). My worldgen tasks were saturating that pool, so the optimize task got queued behind them. block_on_optimize_trees then sat on the main thread waiting for the queue to drain. As more frames piled up without optimize running, each eventual completion was catastrophic.

Flipped use_async_tasks = false:

PhysicsSchedule    total 0.8s (1.6% of update)
optimize          mean 0.79ms / tick, max 7.7ms

Same algorithm, same tree, same data. 45× drop on mean, 95× on max.

tldr; this just moves the work to the main thread synchronously. Cheap for my small static-terrain tree, but a larger / more dynamic tree would eat the cost directly into frame budget instead of (in the happy case) overlapping with simulation. So false isn't a strict upgrade, it just trades latency for predictability. my case benefits because the work itself is sub-ms but the async wait was unbounded under contention.

Default is reasonable when AsyncComputeTaskPool is mostly idle, but any app doing its own bulk compute on that pool hits this silently. Wondering if there's room for a dedicated pool, contention detection, or at least a doc warning?

Versions: avian3d 0.6.1, Bevy 0.18.1, Linux x86_64.

would be happy to get any pointers, suggestions, etc - I am still kinda new to all of these

[hasty rivet]

Why am I gettings this warning? I have all the components in MassPropertiesBundle and the entity has a collider

Collider::circle(16.0),
Mass(10.0),
CenterOfMass,
AngularInertia,

Dynamic rigid body 1298v1 has no mass or inertia. This can cause NaN values. Consider adding a `MassPropertiesBundle` or a `Collider` with mass.

[dreamy carbon]

what's the entity's transform scale?

[hasty rivet]

1.0, 1.0, 1.0

[nimble fiber]

share spawning/mutation code?

[coral kite]

Had a heightfield collider question. The terrain in my game is chunked. I have heightfield colliders for each chunk. I have some real time terrain deformation and currently have to rebuild the whole heightfield collider for the chunk.

Would it be possible for Avian to support partial updates of heightfield colliders instead of replacing the whole collider? (Not trying to demand a feature, just wondering if there are plans to do this)

[cinder summit]

Afaict parry doesn't support this, so unless it is added there, Jondolf has no real control over this (until peck is done, I guess?)

[coral kite]

Damn okay 🙁 guess I could always fork Parry and Avian ><

[vestal minnow]

This is very interesting! Would definitely be useful to mention this in the docs somewhere. If there's some way to detect contention automatically and fall back to the main thread, that could be good to try too.

I believe one of the problems here is that Bevy currently has a 50%/25%/25% split between Compute, AsyncCompute, and IO task pools, which means that any given operation can only be scheduled for a subset of threads. So even if you have lots of free IO or Compute threads, you're limited to just a few AsyncCompute threads, which in your case easily saturates the pool and leads to this contention. There's a PR that aims to fix it, but it's in draft :/
https://github.com/bevyengine/bevy/pull/20699

But yeah in your case I think it makes sense to set use_async_tasks: false. The idea with the async task is to hide the cost of the BVH rebuild/optimization by running it in parallel with the narrow phase and solver where the BVH is not needed. This is also what Box2D does. But the contention obviously makes it not worth it with how the task pools currently work

[vestal minnow]

Yeah like Nise said, I don't think Parry has a way to mutate an existing heightfield other than setting the cell status or some flags related to internal edges. You can't move the actual vertices without reconstructing the heightfield afaik

I don't see a fundamental reason why it couldn't be supported, so I'll probably try to support it in Peck when we implement heightfields for it

[coral kite]

Ooo thanks for the reply

[hasty rivet]

it might be worth noting the warning only appears once when the entity spawns, not every update

#[derive(Component, Reflect, Debug)]
#[reflect(Component)]
#[require(
    Sensor,
    CollidingEntities,
    RigidBody::Dynamic,
    CollisionLayers::player_pickup(),
    Collider::circle(16.0),
    // setting mass to a non zero value (I tested 10.0) does not solve the issue
    Mass::ZERO,
    CenterOfMass,
    AngularInertia,
    Visibility
)]
pub enum Pickupable {...}
// Activated by a chain of observers started in `FixedUpdate`
commands.spawn((
  Name::new("GiftPickup"),
  Sprite::from_atlas_image( atlas.image.clone(), TextureAtlas { layout: atlas.layout.clone(), index: 80, },),
  Transform { translation: transform.translation(), ..default() },
  LockedAxes::ROTATION_LOCKED,
  Pickupable::None,
  CollisionLayers::player_pickup(),
  MagnetizedPickup::default(),
  GiftOnPickup,
  GameEntity,
  // tried adding computed on spawn but it didn't fix the issue
  (ComputedMass::new(1.0), ComputedAngularInertia::new(1.0), ComputedCenterOfMass::new(0.0, 0.0),),
));

[vestal minnow]

You're requiring default AngularInertia here, which I believe is zero. Mass and AngularInertia are only meant to be added (with non-zero values) if you want to override the values that are automatically computed for colliders

You shouldn't need any of those mass property components there unless you specifically want to set them

That body has a collider, so it should get some mass properties automatically

also see docs here for general mass property stuff
https://docs.rs/avian2d/latest/avian2d/dynamics/rigid_body/mass_properties/index.html

I believe another thing that can sometimes cause the warning is if the z component of Transform::scale is 0, that one is a bug

[hasty rivet]

The issue is with the sensor component. Sensor doesn't require any components so why would that happen?

// no warning
#[require(
    // Sensor,
    CollidingEntities,
    CollisionLayers::player_pickup(),
    LinearVelocity::default(),
    Collider::circle(16.0),
    RigidBody::Dynamic,
    Visibility
)]

[nimble fiber]

your component requires default mass properties components so they will still be added.

[vestal minnow]

ah okay so that is this subtle detail in the Sensor docs
https://docs.rs/avian2d/latest/avian2d/collision/collider/struct.Sensor.html

this is consistent with Unity and Godot, though it is a bit footgunny perhaps

So in this case you do need to manually set both Mass and AngularInertia :/ but both need to be set to a non-zero value

[hasty rivet]

That worked, thank you!

[vestal minnow]

you can also use MassPropertiesBundle::from_shape(&collider, 1.0) to derive all of the mass properties from the shape

either way works

[autumn lake]

Is there a reason (or am I doing something wrong) that I can't use &mut CollisionLayers in a query? For example

    mut blocks: Query<(
        Entity,
        &mut MeshMaterial2d<ColorMaterial>,
        &mut CollisionLayers,
    )>,

gives the following error:

type mismatch resolving `<CollisionLayers as Component>::Mutability == Mutable`
required for `&mut avian2d::prelude::CollisionLayers` to implement `QueryData`
1 redundant requirement hidden
required for `(Entity, &mut MeshMaterial2d<ColorMaterial>, &mut CollisionLayers)` to implement `QueryData`

It works if CollisionLayers is not mut, but I need to change it in this system

[nimble fiber]

CollisionLayers was made into an immutable component recently. In order to update it, use a mut commands: Commands system parameter and then commands.entity(your_entity).insert(CollisionLayers::new(...)). You shouldn't need to query for CollisionLayers at all unless you need to reference its previous state.

inserting a new copy will replace the old copy with the new one

[queen hare]

Hey, is there a proper way to take control over the physics stepping of Avian? I want to run the systems of my simulation in lock-step with avian physics and want to be able to pause/resume the simulation.

My current solution feels a bit clumsy:

• pausing the Time<Physics> in the PreStartup
  (Ideally I want to add the PhysicsPlugin/ insert the Time::<Physics> resource such that the physics does not run on its own in any schedule. It should only run "by hand")
• running a "simulation_step" system that advances the Time<Physics> by a fixed delta and running (run_schedule) the PhysicsSchedule and my own SimulationSchedule on the world.

This simulation_step system runs only if the App is in SimulationState::Running.
The problem especially is for some reason I have to initialize the App in the SimulationState::Paused state, otherwise the whole program crashes at the start.

[queen hare]

Okay, I solved the crashing issue, but is there a way to insert the Time::<Physics> resource paused?

[acoustic spoke]

hello, is there a way to have a transforming gravity effect per-entity?

kind of like GravityScale, except it's a Mat3A instead of a f32

i guess there's always the option of removing and replacing the system responsible for adding gravity in 0.19, but i was wondering if this was a planned feature

[vestal minnow]

There's ConstantLinearAcceleration that is basically like a per-entity gravity

"constant" as in applied every frame continuously, you can still mutate it

[acoustic spoke]

oh so i can do like 0 gravity for everything by default and use that component for custom gravity instead?

[vestal minnow]

you could do that yes

[acoustic spoke]

thanks!

[vestal minnow]

np!

[autumn lake]

oh awesome that makes sense, thanks!

[wet niche]

Maybe try

app.insert_resource({
            let mut time = Time::<Physics>::default();
            time.pause();
            time
        })

after adding PhysicsPlugins? If it still crashes, describing the actual crash message/reason would help too.

[queen hare]

Oh, yes. That works. Thank you!
I fixed the crashing. I accidentally ran the PhysicsSchedule twice in a weird way I think. But it runs wonderfully now.

[velvet patio]

yay, https://github.com/dimforge/parry/pull/418 just got merged and parry 0.27.0 just got released that bumps glam, so that solves the dependency mess

i can update my avian branch and open a PR in a bit for 0.19

[sweet sundial]

-# #[cfg(all(feature="2d",feature="3d"))] const _: () = break rust;

[prisma bay]

ngl but doesnt look like its going to get merged anytime soon, too many blockers 💀

[vestal minnow]

yeahhh I don't know what the current plans are there, this PR has been (partially) adopted at least three times now and the original issue is over five years old at this point :P

hopefully something happens there sooner rather than later though

[prisma bay]

side effects of not being a game engine backed by AAA games and rich ass corporate companies ig /s

[acoustic spoke]

compile_error!() no?

or we could just make it trigger a spontaneous UB 💀

[dreamy carbon]

say it's ub and then have it rm -rf /*

[deft marsh]

I'm getting Collisions registered by the Collisions sys param before the colliding entities have actually moved to the point where the collision would be peformed (i.e. their transforms have not registered the new position)

This leads to the objects reacting to the collision before visibly colliding, especially noticeable in really fast objects

How can I work around this?

[nimble fiber]

Easy thing to try first is to enable extrapolation

[deft marsh]

how do i enable it?

do you mean interpolation? if so, it is already enabled

wait, no, it isn't

how do i enable it?

ok i just read the docs

did nothing

just tried interpolation only, extrapolation only, and both. i got the same result