aHiroakiIshikawa/so101_6dof_2cameras_yellow_cube_failure_20261024_180959_negative_open_to_grasp

Open-to-grasp negative segments

Source: aHiroakiIshikawa/so101_6dof_2cameras_yellow_cube_failure_20261024_180959_trimmed, pinned revision e0b8b2c59fd6a840dacac722711aadef8126b929.

ユーザー指定に基づき、各エピソードの最初のgripper開口完了直後から最初の把持動作の閉口・接触相当の停止までを切り出した不例データです。2カメラ映像と実測gripper位置で区間を確認し、その後の持ち上げ・再開口・復帰は含めていません。実測位置の停止は把持成功や接触センサーによる検証を意味しません。

• 5 episodes / 829 frames / 27.633 seconds / 30 fps / 2 cameras / 7D actions
• LeRobot v3.0; original task and action/state values preserved.
• annotation.is_negative = true on every retained row; segment annotations and exact source boundaries in trim_manifest.json.
• Timestamps, indices, video offsets and all statistics recalculated. Videos encoded H.264 CRF 0; every retained decoded YUV frame verified identical to source.
• Start: one frame after measured opening reaches within 0.5 units of initial open plateau (median frames 110–159). End: first frame of six-frame closed plateau (range <0.25 units), after gripper falls >1 unit below open plateau; endpoint included. Frame ranges are zero-based, end-exclusive.

CARF motivates separating failure-critical behaviors. These are user-directed phase-level negative labels, not predictions from CARF's learned progress scorer. They do not establish that every individual action is failure-critical; CARF training still requires an appropriate scorer or an explicitly adapted negative-segment training pipeline.