Public-YAM-runs

Physical bimanual YAM episodes recorded by the BluPe operator station. Each run adds an episode to this repository. Failed, interrupted, stopped and timed-out runs are retained and labeled; these are not all successful demonstrations. A model saying done is not independently verified task success.

Loading

from datasets import load_dataset
runs = load_dataset("andlyu/Public-YAM-runs", split="train")
usable = runs.filter(lambda row: row["training_valid"])

The default configuration preserves the original Image/Parquet archive. The lerobot configuration and meta/ directory provide an additional LeRobot v2.1 layout with separate camera videos. Each atomic commit adds an episode and its metadata. Previously uploaded archive files and replay URLs are retained.

videos/ep_*.mp4 contains a 10 FPS H.264 view arranged left | top | right. Video playback follows the episode timeline, labels stale camera frames, and shows missing-frame gaps rather than silently skipping them. It is a viewing artifact; the original images and timestamps remain in the Parquet rows. The local runner's Save video prefers its browser-recorded replay when available, otherwise this episode's MP4. Save log downloads a separate local ZIP with exact model request/response bodies, model input images, interaction events and errors. Model logs are not uploaded here.

Samples

• episode_id, frame_index, timestamp: episode identity, sample index and seconds relative to the recording start on the Jetson monotonic clock.
• left_image, right_image, top_image: JPEGs of the corresponding camera. The station's capture-time overlay is retained in the pixels.
• measured_joints: 12 measured joint positions, radians, left joints 0–5 followed by right joints 0–5. measured_grippers: left/right normalized positions, 0 closed and 1 open.
• action_joints, action_grippers: the most recent targets actually dispatched by the driver, in the same order/units. They are not substituted for feedback. action_timestamp identifies that dispatch relative to recording start. action_valid=false means no known command; its zero placeholders are not actions.
• Camera timestamps, sequence numbers and frame ages are retained per view; camera_skew_s records inter-camera timing difference.
• training_valid: known action and all three frames at most 150 ms old at the measured sample time. Retained invalid rows should be filtered for training.
• task: task instruction from the API handoff, or null if unavailable.
• outcome: station termination reason; session_api_stop does not distinguish success from a runner/operator stop. session_timeout is timeout, and session_api_disconnect is interrupted communication.

The nominal sampling rate is 10 Hz. USB cameras are software time-aligned, not hardware-triggered; timestamps are host capture-return timestamps rather than sensor exposure timestamps. Do not assume uniform intervals or perfect synchronization. Missing samples and capture errors are counted in the episode manifest. Do not silently interpolate them into measured demonstrations.

Recording ends when policy authority is revoked. Timeout parking and other subsequent recovery motion are excluded from policy samples. Local recordings are retained if publication fails, and the uploader retries independently of arm control. Simulated episodes and model transcripts are not published here.

Provenance and limitations

All recorded poses are robot joint-space state, not calibrated world poses. Camera role mapping is left=10, right=4, top=16 on this station; camera calibration and image-space task success are not supplied. Review outcome, timing validity and dataset quality before training. No success-only filtering is implied by the train split name.

LeRobot v2.1 representation

Canonical files are data/chunk-000/episode_000000.parquet and videos/chunk-000/observation.images.{left,top,right}/episode_000000.mp4, with path templates, dimensions, counts, tasks and statistics in meta/. meta/blupe_source_episodes.jsonl maps stable numeric indices to original episode IDs, source hashes, outcomes and recorder quality counts.

State and action are float32 vectors with this exact order: left joints 0–5, left gripper, right joints 0–5, right gripper. Joint units are radians; grippers are normalized 0–1. State is measured feedback; action is the most recently dispatched target. Unknown actions retain zero placeholders with action_valid=false. A missing action timestamp is represented as zero and must be interpreted using that validity flag.

Timing and training validity

The canonical playback clock is timestamp = frame_index / 10, with exactly one frame per original recorded row in each camera video. This compresses missed acquisition time; it is not the physical elapsed-time clock. Original elapsed seconds remain unchanged in float64 source_timestamp. frame_gap=true marks a row whose interval from the previous recorded row exceeded 0.15 seconds. No images or joint values are interpolated or invented. Use original timestamps and the raw joined replay for physical elapsed-time analysis.

training_valid preserves the recorder's per-row action/camera validity; it does not certify continuous timing. For temporal/action-chunk training, reject or split windows crossing any frame_gap, and reject invalid rows. Simply filtering out gap rows and then treating remaining rows as adjacent would still be incorrect. Use source timestamps if resampling, and keep any synthetic values explicitly invalid. All-frame statistics describe the archive, including invalid action placeholders; recompute normalization statistics after choosing training filters.

from datasets import load_dataset
canonical = load_dataset("andlyu/Public-YAM-runs", "lerobot", split="train")
# Camera features are external MP4s declared by meta/info.json, not Parquet images.

This targets the v2.1 schema and compatible viewers, not the newer v3 writer. No LeRobot package or Torch is required by the Jetson publisher.