I also want to highlight hard work from collaborators to hook up MJX with the Madrona batch renderer (via github.com/shacklettbp/...).

I'm super impressed that Mustafa Haiderbhai and Jing Yuan Luo were able to train a pick-cube policy in 10 minutes from pixels on a 1x RTX 4090

As well as a non-prehensile high frequency torque control policy:

I'm particularly excited about the manipulation results, which used our MuJoCo MJX simulator to train dexterous manipulation policies:

Excited to announce playground.mujoco.org an open-source library for GPU-accelerated robot learning and sim-to-real transfer. With `pip install playground` you get access to a whole host of robotic learning environments, many of which were transferred onto robots!