Recorded robot episodes and labelled failure cases used by our world-model benchmarks. Each dataset note lists provenance, contents, known limitations, and licence terms.
Each note lists the dataset contents, capture setup, licence, and known limitations.
1.4M teleoperated tabletop episodes with paired language instructions, camera intrinsics and per-frame gripper state. The set our language_table cases are drawn from.
Two-arm handover and place-down runs recorded at 60fps from a fixed first-person rig, with contact annotations at every transfer moment.
Generated clips showing object duplication, arm teleportation, identity drift, or physics violations. Each clip is labelled by failed axis.
Multi-minute kitchen sequences: opening, pouring, wiping and stacking chained into single takes for training and testing state persistence.