A humanoid that taught itself to walk. Shove it and watch it catch itself.
A Unitree G1 humanoid in a physics simulation, walked by a small neural network: about 219 thousand numbers. Fifty times a second it reads how the body is moving and sets a target for each of its 29 joints. Nobody wrote the walking. It was learned in simulation, by trial and error.
Drag the robot to push it, or press Shove. Arrow keys, or the stick on a phone, choose where it walks. Pull hard enough and it falls.
Policy by Julien Blanchon, trained with mjlab, from mjswan (Apache 2.0). Robot model and meshes by Unitree Robotics, via MuJoCo Menagerie (BSD 3-Clause). How the policy reads the robot is ported from mjswan by Tatsuki Tsujimoto. Physics: MuJoCo 3.14 (Apache 2.0). Drawing: three.js (MIT). Policy runtime: ONNX Runtime Web (MIT).
Runs on your device. Nothing is uploaded.
Drag the robot to push it · arrow keys steerDrag the robot to push it · the stick steers