Noctis12138/RoboDojo-Aug

RoboDojo augmentation data (cover_blocks, fasten_screws)

Simulated bimanual demonstrations for two RoboDojo tasks, generated (not recorded by the RoboDojo authors) by transforming segments of the released RoboDojo demonstrations to new object layouts and re-planning the motion in the RoboDojo / Isaac Sim environment, then checking every episode with the official RoboDojo success check.

Not official RoboDojo data. Source and environment:

• RoboDojo benchmark: https://github.com/RoboDojo-Benchmark/RoboDojo
• Released RoboDojo demonstrations: https://huggingface.co/datasets/RoboDojo-Benchmark/RoboDojo (Apache-2.0)

No evaluation layouts or evaluation seeds of the benchmark were used to generate this data.

Layout

shards/<task>/<shard>/, each shard about 50 source episodes:

• full/ LeRobot v3.0 dataset: whole episodes, official task instruction, clean images (same format and video encoding as the released RoboDojo LeRobot data: 25 fps, AV1 CRF 18, g 2, yuv420p).
• subtask/ LeRobot v3.0 dataset: one episode per atomic skill (pick, insert, tighten, cover, uncover, hand-over, regrasp, reseat, return_home, ...), sub-task instruction, clean images plus the exact target mask of every camera (observation.masks.<cam>, lossless PNG). The mask overlay used for training is applied at data loading.
• every source episode appears twice: the standard appearance and one random appearance (room, table / floor material, lighting, visual-only distractors); both replay the identical motion.
• about 10 % are corner cases (an injected error and its recovery); the injected segment itself is in no dataset, full episodes of corner cases start where the injection ended.
• aug/episodes.jsonl generation metadata (layout, seeds, source segments, atomic-skill records), aug/commands/*.npz commanded actions and object poses (replaying them reproduces the episode), attempts.jsonl all generation attempts, shard.json machine / GPU / seeds, manifest.json sizes and SHA-256.