ONNX

[beni-ea]

Hi,

The onnx file in this repo is a different than the ones running on the actual robots.

The onnx files on our robot (PF_TRON1A_129) have the following format:
name: mlp_input
tensor: float32[68]
name: mlp_output
tensor: float32[6]

The one in this repo:
name: nn_input
tensor: float32[27]

name: nn_output
tensor: float32[6]

Could you provide source code for the controller running on the robot?

Thanks,
Benjamin

[limxdynamics]

Thanks for the detailed report, and sorry for the slow reply!

You were right — the policy.onnx in this repo at the time was a simplified version (27-dim input), different from the 68-dim model on the actual robot. The model has since been updated (now 36-dim, with a separate encoder.onnx for observation history), though it still differs from the 68-dim production model.

The public model here is meant for dev/sim reference — full production training details aren't open-sourced yet. For closer real-robot behavior, training your own policy via our open pipeline (e.g. pointfoot-rl-isaacgym) is the best option right now.

We'll check internally on the current production model and update here if there's more to share. Thanks again for flagging this!