Hold to accelerate, release to decelerate smoothly back to a stop — press the opposite key or Esc to change it.

Pause/Restart affect the sim run itself, separate from driving it or switching policies. Auto-restart timer is off by default — the robot only resets when you hit Restart (or on a real fall) — turn it on to get the training-style episode timeout back.

What the running robot's state actually is, right now — read these before picking a Create Policy target so you're choosing a real number, not guessing. sensor is a real IMU/gyroscope reading, available on both sim and real hardware. sim only exists in simulation — no equivalent sensor exists on the physical robot, so it can't be an input to a real-robot policy, only a training-time target.

What the robot sees, from a camera positioned roughly where its head-mounted sensor would be — a robot-POV RGB feed, separate from the third-person Simulator view. Hidden unless the server was started with --camera.

Drag any indicator, use W/A/S/D and all four arrow keys (↑/↓ are aliases for W/S), or the mouse-look pad to set the velocity command. +X forward, +Y left, +Yaw turns left.

Building your own controller (gamepad, custom hardware)? This panel sends the exact same WebSocket calls a home-made joystick would — see Talking to the robot: the control protocol in the Docs tab.

Which registered task this whole session is running (e.g. "g1" walking vs "g1_target" standing) — different from picking a Policy below, which only switches WITHIN the current family. Switching family restarts this server's process for the new task (10-20s) — the page reconnects on its own once it's back. Only tasks with at least one trained local policy are shown. If the Simulator view stays frozen once it's back, use the ↻ button to reconnect it manually — this isn't automatic because reloading it is CPU-heavy and can itself slow down the robot's control loop, so it's your call when to pay that cost.

Reference-motion clips this task can track (mjlab's Rugiar-G1-Mimic only — hidden entirely on a backend with no motion-tracking concept, e.g. Genesis). Picking a clip restarts this server's process against it, same 10-20s reconnect as a Family switch above.

Clips marked no policy have no local policy trained against them yet — picking one is NOT disabled: the reference-motion "ghost" overlay plays regardless of what's driving the robot mesh, so you can preview a clip (the robot holding a neutral pose, or whatever policy happens to be active) before ever training anything against it.

Click a policy to switch, or press its shortcut key. Shortcuts are fixed to position, not name — drag ⋮⋮ to reorder: top is 0, then 9 down to 1 (up to 10 policies).

Training runs this policy through random shoves and a velocity command that changes on its own. Turn both off to drive it deliberately instead — like a real operator would, not like a training run.

Trains a new policy asynchronously (a real subprocess, not a UI trick) and adds it to the Policies list above once it finishes. The command below is exactly what will run — nothing is hidden behind this form.