Datasets:

saybeok03
/

orca-hand-bottle

Robot: AM-ARM200 arm (3 DoF) + right ORCA Hand v1 Feetech (17 DoF).
RGB cameras: bottom, wrist and RealSense, each 640x480.
observation.state: measured 20-joint canonical model angles in radians.
action: last successfully dispatched 20-joint target in the same frame.
Format: LeRobot v3.0, generated with LeRobot 0.4.4; MP4 videos and Parquet rows.
Joint order, arm zero/signs and hand conversion convention: robot_metadata.json.
Source-session provenance: collection_manifest.json.

Failed retries were discarded during recording. The initial 75 episodes excluded recordings shorter than 20 seconds. This revision appends all 90 operator-selected recordings from three new sessions, including 13 shorter than 20 seconds, as explicitly requested by the operator. Successful task completion is operator-labelled; no automated success scoring is claimed. Camera synchronization is recorder-side sampling, not hardware-triggered exposure synchronization. Hand angles are derived from motor encoders, not independent link-angle sensors.

Training

Use the dataset repository ID with LeRobot. Data remain 20-dimensional and in canonical model radians: do not convert them to degrees or manually pad saved rows. Pi05 input preparation handles image resizing and dimensional padding. meta/stats.json includes exact numerical q01/q99 over all published rows. For held-out evaluation, fit normalization on the chosen training split only. No held-out performance or physical rollout success rate is reported here.

The train metadata split exposes all 165 episodes. Future collection revisions should be pinned by commit or dataset snapshot when reporting training results.

Collection revision

collection-v2-165 contains 165 episodes and 141,740 frames. The first 75 episodes retain their original indices and video bytes. The previous collection-v1-75 snapshot remains available.

Downloads last month: 131

Models trained or fine-tuned on saybeok03/orca-hand-bottle