PROMO
Go2 Loading t 0.0 s
loading…

PROMO

Loading MuJoCo, policy and model…

Base angular acceleration

Command vs measured vel Dashed = command

Power

Preference changeRobot switch KickReset

PROMO quadruped simulator

A Unitree quadruped (Go1 or Go2) walks with a preference-conditioned reinforcement-learning policy. Move the preference point between tracking, stability and efficiency and watch the gait, torque and power change — live.

WS / ↑↓Forward / backward (hold)
ADStrafe left / right
QE / ←→Turn left / right
SpaceStop (zero command)
K / EnterApply the kick  ·  [] size  ·  ,. direction
1–4Preference presets: Balanced · Tracking · Stability · Efficiency
R P HReset · Pause · This help
Mouse / touchDrag = orbit the camera, wheel / pinch = zoom; the on-screen keys work with touch

The Deploy policy is the one used for the real-hardware PROMO experiments. It was trained on the Go2 in Isaac Lab (PhysX, flat ground, blind, heavily randomised) and runs here in MuJoCo (WebAssembly) with the same delayed DC-motor actuator model. The Go1 is an experiment: the same policy runs zero-shot on a Go1 model (lighter, different geometry and motors). Behaviour can differ from the paper's simulator. Settings beyond the training ranges are flagged out of distribution.