qinghangliu/PAL_realworld

PAL Real-World Evaluation Videos

This dataset publishes real-robot experiment videos for Ours, FrameSamp-Modul, and MemER on PatternLock, PickXtimes, VideoPlaceOrder, and VideoUnmask. It contains all 1,257 MP4 videos from the source evaluation bundle, including 1,170 videos associated with 339 available evaluation records and 87 supplementary backup/demo videos. Multiple camera views and review videos can belong to the same record, so the number of video files is not the number of evaluation episodes.

Download

pip install -U huggingface_hub
hf download qinghangliu/PAL_realworld --repo-type dataset --local-dir PAL_realworld

Download videos for one method/task, for example:

hf download qinghangliu/PAL_realworld --repo-type dataset \
  --include "videos/ours/PatternLock/**" --local-dir PAL_realworld

Individual MP4 files are available under the repository's Files tab. The videos/metadata.jsonl index records each video's method, task, source run ID, outcome annotation, view, duration, dimensions, byte count, and SHA-256. Its file_name paths are relative to the videos/ directory.

Layout

README.md
success_rates.md
release_summary.json
evaluation_records.jsonl
video_checksums.sha256
raw_observation_videos.md
raw_observation_videos.json
videos/
  metadata.jsonl
  ours/<task>/<run_id>/...
  framesamp-modul/<task>/<run_id>/...
  memer/<task>/<run_id>/...
  supplementary/backup/...

evaluation_records.jsonl contains one row for each available evaluated run. Success/failure annotations are preserved from the source run-directory labels (s... / f...); they are not inferred from video filenames or the runtime's unknown, interrupted, or timeout suffixes. The aggregate table below is reproduced from the supplied success_rates.md without changing its values. Supplementary backup/demo videos are kept for completeness and are not additional episodes in that table. File paths within each run retain their source layout.

Real-World Success Rates

Success rates are reported as mean +/- sample standard deviation (ddof=1) over three trials. Each task contains 10 episodes per trial, and missing rollout records are counted as failures in their corresponding trial. The overall rate is the unweighted mean across the four task rates within each trial; its standard deviation is computed across the three trial-level overall rates.

Method	PatternLock	PickXtimes	VideoPlaceOrder	VideoUnmask	Overall
Ours	40.0 +/- 0.0	53.3 +/- 15.3	50.0 +/- 0.0	100.0 +/- 0.0	60.8 +/- 3.8
FrameSamp-Modul	30.0 +/- 10.0	40.0 +/- 30.0	13.3 +/- 5.8	30.0 +/- 0.0	28.3 +/- 3.8
MemER	33.3 +/- 5.8	0.0 +/- 0.0	46.7 +/- 30.6	100.0 +/- 0.0	45.0 +/- 6.6

Missing Records

• Ours/PatternLock: 15 of 30 rollout records are missing and counted as failures.
• Ours/VideoPlaceOrder: 6 of 30 rollout records are missing and counted as failures.
• All other method-task combinations contain 30 rollout records.

The missing rollout records correspond to episodes whose generated subtask list was incorrect. We therefore did not proceed to robot execution for those episodes, so no execution video was produced. These episodes are still counted as failures, and the denominator remains 10 episodes per trial; they are not excluded from the reported success rates.

Video provenance

MP4 files are copied without transcoding; every published copy has been checked against its source using SHA-256. Some records have annotated review videos, some have combined rollout videos, and camera streams are present where they were available in the source bundle. No missing video has been synthesized.

Raw camera videos use fixed 30 FPS in ascending image-frame order, rather than wall-clock action timestamps. The source conversion report records 72 corrupt or truncated PNG frames that were replaced in the camera videos by the nearest preceding decodable frame (or the following frame at the start). This preserves frame indices and stream lengths. Details are retained in raw_observation_videos.md and the report raw_observation_videos.json, with portable relative paths.

Verify the downloaded video bytes from the repository root:

sha256sum -c video_checksums.sha256