[Question] Trained model differs from released parkour_onboard_preview_stair - stair climbing behavior and gap traversal issues

[rllearnerntu]

Description

Thank you for the impressive work! I am currently encountering some issues. I trained a policy using Instinct-Parkour-Target-Amp-G1-v0 following the README instructions, but the trained model behaves differently from the released parkour_onboard_preview_stair model:

1. Standing: My trained model keeps taking small steps even when velocity command = 0, while the released model stands still properly
2. Stair climbing: The released parkour_onboard_preview_stair model climbs stairs very smoothly on the real robot. However, my trained model fails to climb stairs properly.
3. Gap traversal: The released model seems only for stairs and cannot cross gaps at all

Experiment Setup:

• Training command: python scripts/instinct_rl/train.py --headless --task=Instinct-Parkour-Target-Amp-G1-v0 for 30000 iterations
• Hardware: G1 29dof

Any insights would be appreciated!

[ZZQ0316]

Hello friend, did you deploy the released models with instinct_onboard without any change?

[rllearnerntu]

Hello friend, did you deploy the released models with instinct_onboard without any change?

No, the deploy code works well. I didnt change anything.

[zst1406217]

Hello,
As mentioned in #25, we removed some noise models and domain randomization in the training config. Due to hardware differences in real robots, please determine these parameters yourself.
Also, as stated in the paper, some policies are task-specific. The provided stair policy is finetuned only on stairs, so it cannot traverse gaps.

[rllearnerntu]

Got it, thanks for the clarification! Since the G1 29DOF I'm using is the same one used in your paper, which specific noise or randomization parameters do you find most critical for the sim to real transfer of the parkour task? Any recommendations would be a huge help.

[zst1406217]

For perception depth, please refer to our paper. For domain rand, the critical one is randomize_camera_offsets. We provide many noise models and mdp functions in InstinctLab, please try to determine by yourself.