GELLO Bimanual Teleop — 5 Golden Dataset

Five hand-picked, QC-passed episodes of bimanual (two-arm) robot teleoperation captured via a GELLO leader-arm rig. Tasks: handover cube left-to-right, relay chain blocks, share-carry basket handles, two-arm carry tray, and re-grip fragile object for safer carry. Each episode ships with the scene video (1280x720 MP4), a Tier-1 numeric QC report (JSON + Markdown), and a Tier-2 semantic QC report produced by NVIDIA Cosmos3-Nano-Reasoner with self-consistency grounding. Semantic annotations include a Tier-1 whole-episode caption and Tier-2 time-segmented sub-action labels with start/end timestamps, plus per-keyframe visibility/consistency grounding metrics. All 5 episodes are numeric PASS (0 FAIL) and semantic PASS — 2 at high_confidence_pass, 3 flagged needs_human_review. Capture rate 9.7-24.8 fps, episode length 10.7-72.3 s, total ~145 s of two-arm manipulation. Note: camera_intrinsics_present and resolution_meets_1080p are reported as WARNING rather than FAIL, since they are hand/egocentric checks that do not apply to GELLO bimanual robot teleop (no 3D hand reprojection; robot wrist cameras are natively sub-1080p).