sim_ep_03

43 simulated cup-placement episodes derived from the real recordings in mzxuan/manipulation_realworld. Each source target is preserved (16 pink, 14 yellow, 13 blue); clutter is randomized. The reconstructed environment and appearance follow the source setup. Actions come from a state-feedback simulation expert, not extraction from video.

• dataset/local/sim_ep_03_train and sim_ep_03_validation: validated LeRobot training splits.
• raw/episodeXXXX.tar.gz: accepted raw rollout, lossless camera frames, camera videos, synchronized state/action HDF5, scene specification, contact traces, and validation.
• debug/videos and debug/panels: real/simulation comparisons for all 43 episodes.
• inputs/prepared_scenes_and_appearance.tar.gz: real-derived scene inputs and appearance assets.
• reproducibility.tar.gz: frozen configuration, code snapshots, provenance and acceptance reports.
• UPLOAD_MANIFEST.json: sizes and SHA-256 checksums of published files.

Control/capture is 15 Hz, physics/contact logging is 300 Hz, with seven state/action channels and training horizon 50. Cameras are wall, wrist, and body_forward. All 43 episodes, the actual OpenPI loader, and 129 debug videos plus 43 panels passed the completion gate before upload. Consult the packaged reports for individual checks.

Original recordings remain in the linked source dataset. Source archive SHA-256: 85e7410827c2b5d33bc7099b7e6a8833f5fd0331959463c9f24199a289d3f85b. Archives retain paths relative to the generation run; JSON provenance may contain original machine paths that must be remapped when reproducing the simulation elsewhere. Failed attempts, runtime caches, credentials, and dependency installations are not included.