Episodes Preview grabette Visualizer
122 episodes · 50 fps · 1 camera · 1296×972 h264

Grabette DROID-W + IMU, consolidated LeRobot v3

Public dataset: 122 recordings, 84,959 frames at 50 FPS, original 1296 × 972 RGB. Includes 88 previously validated recordings and 34 accepted recordings from the pick-and-place retry.

Recording outcomes

The retry processed all 35 selected right-camera recordings independently. 34 were exported; 1 failed. Full source IDs, names, stages and errors: annotation_studio/recording_audit.csv and annotation_studio/failures.json. Recovery Job: https://huggingface.co/jobs/pepijn223/6ae0974c3fb22a235dd2a1f5.

Rejected recordings: 20261001_182422/right (source 20: imu_fusion).

Rejected calibration/fusion outputs never enter the dataset. Camera intrinsics, DROID-W settings and IMU acceptance thresholds are unchanged. No AprilTags or original trajectories enter inference.

Coordinate frame and timing

Camera poses map the current optical camera to each recording's first optical camera. XYZ is in metres; X right, Y down, Z forward. Recordings use independent origins. Metric accuracy needs independent validation.

36 samples lack IMU support and retain visual estimates; 467 have capture-to-grid timing error over 50 ms; 11814 repeat a captured RGB frame on the uniform output grid. Flags preserve these distinctions.

Lossless aggregation and provenance

Merged with LeRobot aggregation tools, preserving source MP4 bytes and 4×4 pose schemas. Global bookkeeping statistics are recomputed from final rows. All original poses, capture times and calibration sidecars are retained. Gripper targets are derived from measured raw telemetry; see the gripper recovery section.

annotation_studio/episode_map.parquet (also CSV/JSON) links each output episode to its pinned source conversion and raw recording. Original report-internal IDs remain source-local; mapping files give the consolidated IDs. Retry reports are under annotation_studio/retries/pick-and-place-20261027/; historical failed-batch evidence remains under annotation_studio/history/ and annotation_studio/excluded/.

All prior episodes and observations remain unchanged. Source data retains its source repositories' terms; consolidation grants no additional rights.

Recovered gripper targets

All 122 recordings / 84,959 RGB rows include recovered raw gripper angles and derived strategy/closure targets. action is 11D: absolute gripper-centre XYZ in the initial optical camera frame, row-based rotation6D, strategy, closure. observation.state is 2D measured continuous strategy/closure. Grasp periods use the original Grabette full-close projection and latched strategy; these are derived labels, not recorded servo commands. observation.gripper_joint_position retains proximal/distal radians with recorded signs.

observation.gripper_pose and observation.gripper_position use each recording's saved camera-to-gripper-centre transform. Camera poses, RGB files, episode IDs and existing observations are unchanged. The metric-scale limitations above also apply to derived Cartesian gripper positions.

Angles are interpolated at the selected original RGB capture_time_s; no IMU offset or source cable correction is reapplied. Every current row has supported gripper telemetry. Sample-age, interpolation-gap, endpoint-hold and validity flags remain available. Original angle files, aligned values and reports are under each annotation_studio/episodes/<episode>/; annotation_studio/gripper-recovery.json gives the full audit. meta/grasp_projection.json records the 93.5° / 102° calibration. Raw tactile, depth and other inputs remain in the pinned source repositories.

Downloads last month
1,881

Models trained or fine-tuned on pepijn223/grabette-slam-consolidated-20261027