G1 model falls despite low timestep and minimal sim speed – possible planner sync issue?

[yunsooSeo]

Hi, I believe I'm facing a similar issue to the one discussed here(#135 ).

When I run the provided example tasks like humanoid_stand or humanoid_walk, the robots seem to behave reasonably well—even with the default timestep = 0.015—as long as I tune some parameters like balance. I don’t need to reduce the simulation speed to the minimum in those cases.

However, I’m now trying to run a similar setup using my own G1 model, using the same residual and transition structure as in the humanoid tasks. Initially, I was getting rollout divergence errors with very high qacc, so I reduced the planner timestep to 0.001. That eliminated the divergence errors, but now the robot still falls down quickly even with the simulation speed reduced to the minimum.

My question is: how can I tell whether this issue is due to a mismatch between the planner and simulation (e.g. planner being too slow to keep up), or whether it's something wrong with my residual definition or task configuration?

Any suggestions for debugging or confirming the source of the issue would be greatly appreciated. Thanks!

[thowell]

A helpful debugging practice for this scenario is to utilize slowdown in the GUI application: press -.
https://github.com/google-deepmind/mujoco_mpc/blob/main/docs/GUI.md#user-guide

If the task works with slowdown, this would indicated that the hardware is not fast enough for the planning in real time. If the task doesn't work with a small slowdown, for example at 1%, then there might be an issue with the task formulation/configuration.