467
edits
Changes
m
→Add wikilink to MuJoCo (MJX)
====== MuJoCo (MJX)======
Offers a lightweight open-source alternative, supporting maximal coordinate constraints and easier to work with. The MuJoCo_MPC repository, created by Google DeepMind, is a toolset that combines Model Predictive Control (MPC) with the [[MuJoCo ]] physics engine to create real-time behavior synthesis. With the advanced MJX extension, which uses GPU acceleration, it can simulate multiple environments in parallel. One approach is to try to replicate the techniques detailed in the AMP (Adversarial Motion Priors) paper to achieve agile humanoid behavior. For example, implementing a humanoid get-up sequence, which matches what was described in the AMP research.
There’s been collaboration between different projects, like Stompy, to get humanoid simulations up and running. You could try converting Gymnasium to handle the URDF (Universal Robot Description Format) file format. Although converting to MJCF (MuJoCo's XML-based format) may present some challenges, we can still get it to work and refine the motor and actuator setup.
Although [[MuJoCo ]] can be slower in single-environment simulations, the MJX extension and its parallel processing potential make it a solid competitor. Compared to enviroments, like NVIDIA's Isaac Gym, MuJoCo might stand out for its extensibility and rapid development. One goal could be to try recreate the walking, running, and getting-up behaviors described in the AMP paper and use them as a foundation for training robust humanoid movements in simulation.
====== VSim ======
