One policy. Three ways to move.

Choose an intent. Explore a connected world of stairs, slopes, and rough ground.

Seven terrain families · one connected map

P-BFM

Take the controls.

The real policy and simulation run in your browser.
Initial download: approximately 200 MB.

No download until you start.

Follow a motion

Follow a time-varying motion reference while adapting to the terrain. The reference preview plays in sync on flat ground.

TRY AN EXAMPLE

Examples reset the scene and wait for Play.

Reference motion

Full LaFAN recordings and additional SOMA unseen motions.

Import your motion

Retargeted G1 29-joint motion · CSV / JSON · up to 120 s.

Motion file

CSV sample rate (Hz)

CSV quaternion

Format help

Set the CSV sample rate and quaternion order before importing. JSON includes its own metadata. Human motion must first be retargeted to G1. Your file stays in this browser.

Format guide · Example JSON

Starting terrain

Difficulty row

Set speed and heading

Horizontal speed 0.0 m/sWorld heading 0°

Ready to load.

Help

Terrain: Changing the start resets the robot. The entire map stays connected.

Speed and heading: Starts from the current reference commands. Speed scales the captured movement; heading rotates it. Turn gradually.

Main robot: live policy inference and physics.

A fixed P-BFM checkpoint runs at 50 control steps per simulated second with MuJoCo physics at 200 Hz. Slower devices play more slowly without skipping physics. Tracking uses reference motions with behavior latents from the same model's backward encoder. Built-in motions are encoded in advance; imported motions are encoded locally. Goal reaching and reward inference use exported pose and reward presets. A single connected map includes all seven training terrain families. Nominal camera placement and deterministic resets are used; this browser port is not an evaluation benchmark.

While playing, hold Shift and left-drag a body part to pull it toward the mouse. The orange line shows the pull; force is capped at 150 N. Release the mouse or Shift to stop. Pause and reset also release the robot. Push applies a horizontal 100 N force along world +Y for 0.2 simulated seconds. The orange arrow marks its direction; while paused, the push waits for Play. The reference preview stays centered on a flat floor and plays at the same reference time as the main simulation. It shows the reference pose, not another policy rollout. Goal reaching refers to a body pose, not navigation to a map location. If the robot falls, it keeps running until you pause or reset.

MuJoCo WebAssembly · ONNX Runtime · Three.js. Third-party notices.