Depth-conditioned RL · Live MuJoCo
Pretrained depth-conditioned robot policy controlling a Unitree G1 in live MuJoCo physics. The policy was developed by the Perceptive Humanoid Parkour researchers. This website adapts their released browser demonstration; no additional policy training was performed.
A depth encoder feeds a 32-dimensional latent into the 140-dimensional student input. Its 29 joint actions drive the robot at 50 Hz, with 500 Hz physics and 10 Hz depth, in simulated time.
Hold W/S to move, A/D for ±45°, Q/E for ±90°. Y switches speed; Space pauses; Backspace resets. Drag to orbit, pinch or scroll to zoom. Shift + mouse drag on the robot applies a physics force.