Tomb raider 7

[cosmic marsh]

@halcyon lion @vital beacon

[vital beacon]

https://tenor.com/view/ah-shit-here-we-go-again-ah-shit-cj-gta-gta-san-andreas-gif-13933485

[cunning lichen]

So this is what @halcyon lion looks like

[cunning lichen]

I'm with warren though, TR7 has been great for us R4Ps, just long kind of sucks

[crude ether]

tr10 is betterrrrrrrrrrrrrrr

tr7 has the wheel moving back and forht a little on the highway

[cosmic marsh]

Just wait till @halcyon lion sees this

[vital beacon]

I feel you on this

[vital beacon]

https://tenor.com/view/pray-for-you-pray-help-sign-of-the-cross-terrified-gif-17006601

[crude ether]

its present both using torque & angle:(

and long is worse compared to tr10

much, much worse

[vital beacon]

Yes tr7 🚮

[halcyon lion]

TR7 is super jerky and terrible on my Toyota

... until I turn on NNLC with the fine-tuned model then it becomes smooth, perfectly centered in lane and confident
I swear Comma has no idea how good they accidentally made it.

[cunning lichen]

yeah NNLC improves it a lot

[vital beacon]

You shouldn’t have to rely on nnlc

[cunning lichen]

Without NNLC, the steering is pretty bad with R4P

[vital beacon]

Alright so it’s bad.

lol

[halcyon lion]

none of the models can stay in lane without NNLC

[cosmic marsh]

Uhhhhh what? NNLC is amazing but I wouldn’t say thattt

Although Cookie Monster needs nnlc fr

[halcyon lion]

I dunno. Maybe this very-short-list of community supported pariah vehicles that Comma refuses to review/help means the steering control is missing obvious tuning.

OTOH the ordinary RAV4 NNLC tunings almost work on RAV4 Prime so it isn't that different. It's mainly a huge weight and tire difference.

[cosmic marsh]

That makes sense. You keep talking nnlc and I’m going to give you an NNLC Apprentice role 😆

[halcyon lion]

I'm hitting random buttons until something works. It's a freak accident we discovered TR7 + NNLC.

BTW the same thing happened to Sienna's new NNLC model + TR7.

One positive report of Ioniq 6 NNLC + TR7 too

[halcyon lion]

The RAV4_TSS2.json fuzzy model exhibited weaker torque at certain speeds than NNLC disabled. That is a classic sign that the model requires retraining which I did. That's when problems like lane bias went away. Then TR3, TR6 and TR7 suddenly were amazing.

[cosmic marsh]

You asked for it

Look at your roles @halcyon lion

[halcyon lion]

oh no

https://blog.comma.ai/096release/
Comma's blog entry about how they tuned Chevy Bolt with a feed-forward algorithm here is interesting. They mention how things like twilsonco's NN is unscalable because it requires too much actual car testing and can't be simulated.
They implemented a simulator for the Bolt's FF that got very close to the real car. (But subsequently they didn't do this for any other cars.)

Especially interesting: There are still outstanding issues, however. For example, the simulator does not model high-frequency oscillations correctly as they are mostly absent from training data (users do not keep openpilot engaged during oscillations, usually).

Notably TR is notorious for these oscillations. There is an oscillation analysis somewhere (I can't find the link anymore?)
My wild ass guess is TR7 is accidentally awesome with NNLC because of this unintended effect. Somehow.

[cosmic marsh]

You know what that makes sense. And it’s sad the guy who joined comma made that NNFF fit for bolt that shocked the community then left the company very shortly afterwards

[halcyon lion]

If my wild ass guess theory is correct, the reason why nothing after TR7 sticks in lane and feels so confident is because they fixed the oscillation bug.
||DTRv3 perhaps is the closest thing. It stays in lane but meanders a lot.|| <--- I need to retest this.

[cosmic marsh]

Hmmm you may be on to something

[halcyon lion]

Any idea where that oscillation analysis is? How do I repeat that oscillation analysis myself?

@cunning lichen if you have a chance please see how good/bad is DTRv3

[cosmic marsh]

Is it here?

https://commaai.github.io/model_reports/

[halcyon lion]

At some point kumar sent me this and wrote "tr is trash". I don't know where he got it from.
He wrote NNLC is dampening the oscillation. Well OK. Without NNLC many models have bias problems on the left or the right (then frequently cross that line).
What if the act of dampening the oscillation is what makes it awesome?

[vital beacon]

Chill chilll that was tr5-7

[halcyon lion]

perfectly aligning with my crazy ass theory

[vital beacon]

https://tenor.com/view/nope-nah-gif-27078288

[halcyon lion]

NNLC is canceling out most of the left/right oscillation without steering at all.

twilsonco wrote: "That was about the "lookahead" lateral jerk that I use in NNLC, which, essentially, only acts on planned lateral jerk if the future planned lateral jerk sign doesn't change. On straights, for example, there's tons of little corrections that all imply some amount of lateral jerk, but you don't need to overcome the steering friction for such small angle changes. Only larger angle changes require that proactive friction compensation be used. Such larger angle changes, such as when entering/exiting a curve, occur over a couple seconds, and over that time the sign of lateral jerk is constant. So the method effectively ignores lateral jerk on straights and other instances where it's short-lived."

Isn't this quite different from what SP normally does? This appears to be doing something unexpected with TR7.

[vital beacon]

@halcyon lion

vs sgo aka tr13

[halcyon lion]

where did you find this? I want to see the equivalent for Aggressive

[cosmic marsh]

It’s on the comma report website thingy

One sec

https://commaai.github.io/model_reports/secretgoodopenpilot/

They added more branches and this one is tr13^