search

google-deepmind / mujoco_menagerie

A collection of high-quality models for the MuJoCo physics engine, curated by Google DeepMind.
Other
1.25k stars 173 forks source link

Boston Dynamics Spot - unexpected behaviour in MJX #69

Open goncalog opened 2 months ago

goncalog commented 2 months ago

Which model is the issue affecting? Boston Dynamics Spot

What is the issue? Adding the Spot's model to a MJX notebook and rendering a short rollout (e.g. with state = jit_reset(jax.random.PRNGKey(0)) and state = jit_step(state, ctrl)) results in a very unexpected behaviour better described by the first 3 frames (the video remains black for the remainder of the rollout): Frame 1 -

Screenshot 2024-07-04 at 19 38 53

Frame 2 -

Screenshot 2024-07-04 at 19 39 10

Frame 3 -

Screenshot 2024-07-04 at 19 39 33

Is there any additional context you can provide (e.g., a spec sheet or a URDF to show a value mismatch)? It's likely this is me doing something wrong rather than a bug but I had used the exact same setup with the humanoid, robotis op3, and unitree h1 models, and this had never happened.

Any ideas of what may be causing it or what i may try to solve it? Thank you.

scott-yj-yang commented 1 month ago

Hi goncalog,

I have this similar issue when using the CPU version of the mujoco xml file in MJX, and the rendering seems like the agent is falling through the ground. A temporary fix we found is add solimplimit="0 .99 .01" to each of the joint attribute, and it fixed the contact issue in MJX. However, I am not completely sure whether there is any unintended side effect of this change. It will be really helpful if there is a guide of how to convert CPU MuJoCo XML file into a MJX compatible formats.

Scott

yuvaltassa commented 1 month ago

cc @erikfrey

kevinzakka commented 4 weeks ago

Out of curiosity, did you remove the freejoint in the base?