rubatotree/classify-blocks-2-512

classify-blocks-2 512

A LeRobot dataset of 512 simulated single-block classification demonstrations. Each episode contains exactly one loose block, an 18-compartment organizer, a randomized class roster, randomized exemplar layouts, randomized appearance, and a deterministic pick-place trajectory. Structured per-frame state is provided under sidecars/; no HUD preview videos are included.

Each episode defines its own class table covering all existing class ids. The table records the rendered albedo, metallic flag, transparency flag, and one to three surface patterns, and the same values are repeated as target truth on every frame of the episode.

The dataset is intended for experimental imitation-learning and auxiliary-supervision work. It does not establish hardware grasp success, real contact behavior, or a real-world success rate. The per-episode flags remain smoke_only=true and training_eligible=false; the aggregate is explicitly admitted for experimental training by the local training script.

Layout

The root follows the LeRobot dataset layout: meta/, data/, and videos/observation.images.front/. Additional structured 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, aggregate validation, and index metadata.

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

All 512 episodes are Cycles renders from the frozen classify-blocks scene. The trajectories are deterministic and the plant removes the measured random residual. The dataset therefore tests scene variation, class matching, export, annotation, and training plumbing; it does not claim physical grasp reliability.