Environment Dimension Mismatch Error (Expected 37, Got 25)

[xkl-cl]

When running the evaluation script, I encountered two sequential errors:
1.Initial Error:
File "for_robosuite/evaluate_robosuite.py", line 252, in eval_bc
observation = format_observations(obs, ft, camera_names)
TypeError: format_observations() takes 3 positional arguments but 4 were given
2.After fixing the parameter count mismatch, I now encounter an environment dimension error:
Environment expects input dimension: 37
Actual input dimension: 25
Dimension mismatch prevents further execution

[cambel]

Hi @xkl-cl
Could you share more details of what you are trying to do?
it would be helpful if you also share the policy that you are trying to evaluate.

[xkl-cl]

Hi @cambel, thanks for your reply!

According to the README, the project mainly describes right-arm control, while the full task requires the left arm to grasp the hammer first, and then the right arm to wipe the dirt.
For data collection, my process was as follows:
I first recorded demonstrations where the left arm performs the hammer-grasping motion.
Based on these recorded left-arm trajectories, I then recorded right-arm wiping demonstrations.
The final dataset therefore contains demonstrations where the left arm is grasping the hammer and the right arm executes the wiping behavior.
I then trained the policy using this combined dataset and ran the provided evaluation script for validation.
I would like to confirm whether this sequential data collection and combination process is correct,

[bai-520]

Hi @cambel,

Thank you for your response. The policy file is quite large and exceeds the GitHub file size limit for uploads in issues. Would it be possible for me to send it to you via email? If so, please let me know the appropriate email address to use.

[cambel]

@xkl-cl

I see, please try the following:
change the compliance_mode parameter in the policy config (the one in the rollouts folder) from variable_stiffness_full to variable_stiffness_diag_only.

[bai-520]

Hi @cambel,
The issue has been resolved—thank you very much for your help. I have a question regarding dataset creation: originally, the code only allowed the right arm to be controlled after the left arm had fully grasped the hammer. I modified it so that the right arm can be controlled while the left arm is still moving. During training, the terminal reports very low errors, but during validation, the right arm moves toward the wiping position before the left arm finishes grasping, causing collisions with the hammer. Could this validation behavior be due to my modification?

[cambel]

during validation, the right arm moves toward the wiping position before the left arm finishes grasping, causing collisions with the hammer.

It may depend on the number of demonstration and how close both arms are during the demonstration.
It has been a while so I don't remember the details but I think we also collected demonstrations of the right arm while the left is still moving. During validation, the right arm some times move to early and collides with the hammer but not enough to make the task fail.

I did notice that the dynamics of the controller felt very different than what we tried before which made it hard to properly control the robots.
I would suggest playing around with the controller parameters until achieving a realistic smooth motion.
For example, by changing the inner controller kp from 2500 to 500 made it already much better.
on this file: https://github.com/omron-sinicx/CompACT/blob/main/config/sim_bimanual_wiping_compliance.yaml