classify-blocks-2-1-full

A growing LeRobot dataset of simulated single-block classification demonstrations. It starts from the frozen 512-episode classify-blocks-2-1 hardware-geometry revision and appends quota-balanced episodes from the same frozen plan generator until it holds 4096. The same Hub repository is updated in place at 1024, 2048, and 4096 episodes; this revision currently holds 2510 published episodes.

Episodes up to index 511 are copied verbatim from classify-blocks-2-1. Episodes after that share its scene, class table, appearance randomization, motion contract, and export schema; the only build difference is that they are published without the separate review HUD video, which the public aggregate omits anyway.

Motion Contract

• approach_source holds at an 85 mm source hover with bounded lateral and height error;
• align_and_grasp descends through a 25 mm intermediate pose before table contact;
• align_target_slot stops 50 mm above the destination and lower_and_release does not descend further;
• retreat_and_reset closes the gripper and returns the arm to the episode's initial TCP position and full six-joint configuration within the recorded tolerance.

Layout

The root follows the LeRobot dataset layout: meta/, data/, and videos/observation.images.front/. Additional review artifacts are:

• sidecars/<episode>/task_metadata.json: class materials, slots, legal target, and fact rules.
• sidecars/<episode>/frame_states.parquet: one row per exported frame.
• sidecars/<episode>/episode_receipt.json: immutable build receipt and file hashes.
• provenance/: plan, config, lineage, and aggregate validation.

The public aggregate deliberately omits the separate HUD preview videos. The LeRobot front camera video remains the authoritative policy observation.

Interface

The policy-facing observation is one 640x480 front image plus the six joint state values. The action is the absolute six-joint command at 15 Hz. training.action_valid is a supervision mask, not a policy input.

Scope

Every episode is a Cycles render from the frozen classify-blocks scene. The trajectories are deterministic and the plant removes the measured random residual. This dataset tests scene variation, class matching, export, annotation, and training plumbing; it does not claim physical grasp reliability or hardware success.

Last aggregated: 2026-09-30T15:25:14+0800.

Known Exceptions

provenance/dataset_validation.json records this revision as "passed": false on two episodes. Both fail the same organizer-inspection band check and nothing else: in both, all 18 slot corners are inside the wrist-camera image, but one compartment corner lands just past the organizer_inspection_top_fraction = 0.6 boundary.

episode	slots in the top band	worst slot corner y	overshoot	pixels at 480
episode_001712	17/18	0.600613	0.000613	0.29
episode_001874	17/18	0.600149	0.000149	0.07

The check is a hard threshold on a continuous projection, so both are boundary cases rather than framing defects. They are published unchanged and are named in provenance/dataset_validation.json and in dataset/classify-blocks-2-1/validation_2510.json. The in-design repair is a supervisor retry (build_dataset_episode.py --attempt 2, which keeps the stratum and re-draws the roster and geometry seeds) once a render window is available.