Starpilot collects manipulation data in working factories, garages, warehouses and repair shops, captured with four embodiments: head-mounted ego video, ego with 3D hand pose, handheld parallel-jaw grippers, and multi-embodiment grippers with a mobile UMI and on-body SLAM. Every rig records the same core signals: camera views, end-effector pose and gripper state.
Same operators, same industrial settings, different sensors. Pick the one that matches your robot, or train across all of them.
Every rig records time-stamped video, inertial data and pose, expressed in end-effector or camera frames, so episodes from different embodiments can be mixed in one training run.
Clips were picked from real shop floors, not lab tables. Click a setting to open it.
The mobile UMI needs no motion-capture room and no base station. A head-mounted 175° stereo camera runs visual odometry and builds a coloured point map while the operator works. Both grippers are localised in the same world frame from ArUco marker cubes and wrist IMUs.
Because the gripper rigs record metric 6-DoF end-effector poses, a human episode can be replayed on a robot. Below, a tracked handheld-gripper sorting demo is retargeted onto a bimanual YAM arm with inverse kinematics, then re-staged into new environments with a video world model.
Everything in this sample is on Hugging Face. Grab a whole dataset, or open any episode in the explorer and hit Download to get just that clip or segment.
The Starpilot Data Sample is a small, hand-picked public slice. The full corpus spans far more industries, tasks, sites and hours, across every embodiment shown here, and we collect to order.