This work wouldn’t have been possible without the most wonderful cast of collaborators from Berkeley, U. Toronto, Cambridge, Stanford and @deepmind.google.web.brid.gy.

You can get started today
🌐 Project page: playground.mujoco.org
📂 Code: github.com/google-deepm...
📝 Paper: playground.mujoco.org/assets/playg...

playground.mujoco.org integrates high throughput batch rendering on GPU with Madrona-MJX. Using this, we trained a pixel-based block picking policy within minutes - achieving robust 100% success in real world trials.

With the Franka arm and playground.mujoco.org, we trained torque-controlled policies for contact-rich tasks like block reorientation. We obtain smooth, compliant and reactive behaviors running at 200Hz.

Locomotion is important, but what about the finer things in robotics? With playground.mujoco.org we took on the challenging task of dexterous in-hand reorientation with the open-source and low-cos LEAP hand. Our policy below was trained in 30 minutes on 2 4090s.

playground.mujoco.org allows humanoid robots to learn robust walking on different surfaces (below: Berkeley Humanoid, see other humanoid robots on the website).

playground.mujoco.org makes training locomotion policies straightforward and fast:
Quadrupeds like Go1: joystick control, fall recovery, handstands and footstands trained in minutes.

The key to fast progress in robotics? Rapid iteration loops.
playground.mujoco.org streamlines the entire process from simulation to deployment and back.
In just 8 weeks, this unlocked:
- Zero-shot transfer across 5 robots.
- Locomotion, manipulation, and vision-based control.