Simulation
A simulated SO-101. A hosted policy. Move objects in the scene and watch the returned actions drive the arm.
Explore the scene without spending credits. Run sends repeated calls until you stop or the session ends. Call cost unavailable — check model status before running.
Sign in to run a policy · or enter an API key below. Signing in opens your dashboard; return here via Simulation.
Drag objects to move them. These controls do not move a physical robot.
| shoulder_pan | 0.0 | |
|---|---|---|
| shoulder_lift | 0.0 | |
| elbow_flex | 0.0 | |
| wrist_flex | 0.0 | |
| wrist_roll | 0.0 | |
| gripper | 0.0 |
What is real, and what is simulated?
The arm, objects and camera views are simulated in MuJoCo. The model runs remotely and returns real action chunks; every returned action uses one token. Stop ends the session. No physical robot is connected.
ACT and SmolVLA are specific pick-and-place checkpoints, not general-purpose robot controllers. SmolVLA was trained on oranges and a plate in simulation. Editing the task does not guarantee the model will follow it, and these worlds differ from its training scene.
The generated room backdrop is for your view only. The policy sees rendered geometry. Session recordings retain the observations and returned actions for inspection.