madokalif/widowx-pick-v1-augmented

Robustness condition: the same source demonstrations plus deterministic image-only variants (blur, gaussian noise, colour jitter, cutout; seed 2026). States, actions and task text are identical to the clean condition — only the pixels differ.

Hand-guided (kinesthetic) demonstrations on a Trossen WidowX AI 7-DoF arm. The operator moved the arm by hand under firmware gravity compensation (Mode.external_effort with zero desired external effort) while the gripper stayed under position control, so each recorded action is the measured joint pose the arm was guided to, not a commanded target.

Format

LeRobot v3 dataset.

field	shape	units
observation.state	(7,)	6 arm joints in rad, gripper carriage in m
action	(7,)	same, measured (kinesthetic teaching)
observation.images.external	640x480x3	RGB, wrist-mounted RealSense D405

• Task instruction: "pick up the object and place on the hole" (identical for every episode)
• Sample rate: 10 Hz
• Episodes: 220

Collection notes

Every episode passed the collection quality gate: no stale camera frames, no repeated frames, contiguous sampling, joint values inside the arm's limits. Episodes that were marked failure/discarded, or whose sampling gap exceeded the threshold, are excluded.

Recorded with holi-lab/WidowXAI.