MuJoCo engine support

[josephrocca]

DeepMind bought the company behind MuJoCo a while ago and open sourced it. It's a very robust engine with a huge ecosystem around it. It's a general-purpose engine, but it's particularly well-suited to bringing embodied agents that are trained via reinforcement learning into games, since most software for training RL agents uses MuJoCo as the physics engine. The idea is that you train with Python, and then you can deploy on the web with no "domain gap" problem.

Despite being more robust / 'scientifically accurate' than game-oriented physics engines, it is still fast/efficient. Here's an existing wasm port:

• https://github.com/zalo/mujoco_wasm
• https://zalo.github.io/mujoco_wasm/
• https://kzakka.com/robopianist/

I'm not sure how hard this would be to add to Phy, but if it's feasible then that would be very exciting!

Some more links to check out:

• https://gymnasium.farama.org/environments/mujoco/
• https://sites.google.com/view/myosuite/myosim
• https://github.com/MyoHub/myosuite
• https://medium.com/embarkstudios/transforming-animation-with-machine-learning-27ac694590c
• https://github.com/google-deepmind/mujoco

[lo-th]

yes is a very good project but is very complex to include, and maybe too specific. there are no javascript function to make bridge with other engine.

i already call them, if is possible to have script function to command directly but ...? not sure is possible

[josephrocca]

Deformable and non-convex bodies with multi-threading :drooling_face: https://github.com/google-deepmind/mujoco/discussions/1101