How this works

• The cell is MuJoCo. The arms, table, mat, block and bin are the compiled twin model the training data was rendered from. In Drive, MuJoCo itself runs here (WebAssembly, 1 ms steps): you set the arms' motor targets, and gravity, contacts and the grasp do the rest.
• Policy uses the pinned block-in-bin π0.5 checkpoint and all three learned training camera renderers through a local simulation bridge. It replans every eight targets at 30 Hz, pausing simulated time during inference. The camera input is either the learned training renders or, with MuJoCo direct, this page's own simulator frames from the same three calibrated cameras (always light theme). The three policy thumbnails show its exact latest input. The v1 preview is never a policy input.
• The table is real. Its texture is the overhead camera's photo of the empty table, assembled from hundreds of frames with the arms masked out.
• The cameras are the rig's. Overhead is the registered overhead camera; the wrist views ride on the grippers at the mounts calibrated from the wrist footage.
• Rendered is a learned model (v1, in this tab) turning MuJoCo's overhead view into a real-looking frame. Replay shows held-out recordings next to the real camera.

Physics is not tuned to the real arms' dynamics. The v1 renderer only knows this cell, and poses far from the recordings render poorly.