UMI cup pickup pilot
One paired handheld UMI demonstration of pick up cup, recorded on October 19, 2026. This is a reviewed pilot recording with documented annotation limitations, not a fully qualified training dataset.
- 1 episode, 219 frames, 7.3 seconds at 30 FPS.
- Two 800 x 600 camera videos, 14D state, 12D pose-delta actions, and a 16 x 32 tactile array on UMI1.
- The earlier setup/button tests are excluded.
- Watch UMI1 | Watch UMI2
- Review measurements and limitations
Processing repair
The original UMI1 reconstruction diverged by hundreds of metres. Its first decoded LiDAR cloud contained only 729 points over 1.096 ms. Removing exactly that startup fragment, while preserving all other serialized sensor messages, their order and timestamps, recovered plausible motion using the same GLIM ROBUST settings. Cross-arm alignment and the LeRobot dataset were then rebuilt. The repaired end-effector tracks have maximum excursions of approximately 20.7 cm (UMI1) and 32.8 cm (UMI2); this is a plausibility check, not a ground-truth accuracy measurement. Raw recordings remain unchanged.
Known limitations
- Pose estimates hold their first available value through the initial 1.40 s on UMI1 and 3.40 s on UMI2. Both streams also hold their final pose for about 0.37 s. Motion in those intervals is not independently estimated. Both tracks overlap on output frames 102-206 inclusive.
- UMI2 gripper-tag detection has a 3.664 s gap, with gripper widths interpolated across video time 0.570-4.234 s. Its camera view is dark.
- Cross-arm alignment is automatic and coarse: 25 cm map voxels and a 50 cm matching threshold. The forward registration RMSE is 18.2 cm under that threshold; this is not a direct gripper-position error measurement or independent calibration validation.
- Tactile values are raw ADC readings, not calibrated forces. Timestamps are host serial-receipt times. All output frames have a valid tactile sample under the pipeline's age gate.
- State includes normalized gripper widths. Actions include pose deltas only and do not include gripper actions.
Use the quality report when choosing frames or interpreting motion. The original failed reconstruction and earlier setup tests are kept locally and are not included here.
Contents and provenance
The meta/, data/, and videos/ directories use the LeRobot v3 format. The raw_episodes/paired_sess_1792443130854323403/ directory contains the paired original camera, LiDAR, tactile and calibration payloads. Processing is based on AfterQuery UMI-Live, with tactile compatibility and the startup-scan repair described above.
The dataset is public under Apache 2.0, selected by the dataset owner.
- Downloads last month
- 44