MLAgents Inference Behavior not matching Learning Behavior

[restive ivy]

Hey y'all, I'm experiencing a weird problem. I have some onnx models that I've produced using mlagents-learn, and now I'm trying to load them for inference in a scene.

When I define an agent and load that model into its behavior through the editor GUI, the behavior during inference matches what I saw during training.

When I instantiate the agent through code, though, and then use setmodel, its behavior becomes nonsensical. Has anyone seen this before, or does anyone have examples of instantiating an agent and using setmodel?

I've yet to try setmodel on an agent that I created through the editor, but I'm gonna do that to try to figure out if the issue is with the agent being instaniated in code, or if it's with setmodel.

These dudes are loading a model that was running laps correctly during training, but during inference they go nuts

I'm also using the exact same settings for inference, same character, same scripts, etc

[restive ivy]

Ugh it was just me being dumb, and I was training on a modified version of the prefab. Eyes were slightly offset, different drag. Fixed that and it worked great

Although it seems that timescale affects training vs inference performance. The ones that I trained at timescale 20 were trash on timescale 1