How to log the rollout of the mujoco sim?

[jaraujo98]

I'd like to compute some tracking metrics on the mujoco rollout of the policy. I tried logging the contents of /joint_states using

ros2 topic echo --full_length /joint_states

but that does not give me the root transform.

I also tried writing a script that runs inference in MuJoCo directly (without using ROS), but without success. Does your first observation in IsaacLab and in MuJoCo match exactly? Mine has some slight differences (on the order of 0.01), but I would guess domain randomization would make the policy robust to those fluctuations.

I also noticed that your ROS MuJoCo runs at a time step of 0.002, while IsaacLab runs at 0.005, so I suppose that there is more to getting the policy to work in MuJoCo than simply matching the joint indices between MuJoCo and the order that the policy expects (plus the parameters like action scale, offset, stiffness, and damping). For sim2sim, is the following what your system is doing:

1. MuJoCo is running at 500Hz
2. There's a separate controller also running at 500Hz that calculates the torques based on the latest action and sends them to MuJoCo
3. The policy is running at 50Hz, setting the PD targets of the controller

?

[qiayuanl]

MuJoCo serves as a sim-to-sim process that we aim to align with the real system as closely as possible. Running it with smaller time steps, incorporating the state estimator, and using the same hardware interface as on the robot are all efforts we take to maximize fidelity.

On the other hand, for successful sim-to-real transfer, the deployment pipeline must remain consistent with the RL training environment—this is exactly what the three bullet points you mentioned ensure.

When it comes to evaluating policy tracking error, it may be more appropriate to measure it at the controller level.

[jaraujo98]

Thank you for the explanation. Could you please elaborate on what you mean with measuring tracking error at the controller level? I would expect the controller to receive PD targets and output torques, so I don't see how I could get the information I need to compute (for example) the global per joint link position error (as is done in HuB). Is there a controller specific topic or service I should subscribe to?