Dataset Viewer
Auto-converted to Parquet Duplicate
Search is not available for this dataset
accepted
bool
checked_videos
int64
counts
dict
errors
list
true
105
{ "baseline": 5, "clutter": 50, "obstacle": 50 }
[]

Isaac Sim from Real

110 MP4 videos: 105 simulated videos and 5 corresponding real recordings from episodes 0, 10, 16, 27, and 36 of the October 20, 2026 UR5e cup-placement dataset.

The simulated set contains 5 baseline replays, 50 clutter augmentations, and 50 obstacle augmentations. Each episode has one baseline, ten clutter variants, and ten obstacle variants. Total MP4 size: 624.7 MB.

Contents

  • real/episodeNNNN.mp4: the corresponding real recording, with wall, wrist, and body-forward views arranged left to right (1920 × 480, 15 fps). All source frames are included. These are viewing exports from the original recordings, not new real-world executions of the augmented tasks.
  • real/episodeNNNN.timestamps.csv: original frame timing; MP4 playback uses the nominal 15 fps recording rate.
  • simulated/episodeNNNN/baseline.mp4: the recorded robot commands replayed in simulation.
  • simulated/episodeNNNN/clutter_00.mp4 through clutter_09.mp4: added tabletop objects with the original recorded commands.
  • simulated/episodeNNNN/obstacle_00.mp4 through obstacle_09.mp4: obstacles requiring changed behavior. Across the set, 17 use detours and 33 clear an obstruction before continuing.
  • Adjacent JPG files show initial, intermediate, and final frames. Simulated MP4s are 640 × 480 at 15 fps.
  • manifest.json and videos.csv: file inventory, real/simulation pairing, scenarios, durations, sizes, and SHA-256 hashes.
  • SHA256SUMS: file checksums.

Appearance and provenance

Robot and task objects are rendered 3D models with simulated motion and physical contacts. Room/cart backgrounds use static photographs from the real recordings mapped onto approximate surfaces. The table texture combines observed wood from recordings of the same workcell. The background is not a complete 3D room reconstruction, and novel viewpoints can expose projection errors. The real videos and simulated videos are related source/replay examples, not guaranteed frame-aligned pairs. These five recordings were used for reconstruction and replay; no held-out test split is claimed.

The simulated videos are the completed augmentation_105_realistic_v2_20261022 batch. Its stored automated audit reports 105 accepted videos. Real exports were fully decoded and their frame counts checked against the original HDF5 recordings before upload.

Textured prop sources: Poly Haven. Individual source URLs and source asset licenses are listed in asset_credits.json. Those asset licenses do not assign a license to the real recordings or the entire dataset.

Downloads last month
65