RoboFAPE Dataset v1.0.0

This dataset supports robot progress estimation, failure detection and task-conditioned trajectory filtering. It accompanies RoboFAPE: Failure-Aware Progress Estimation for Robot Manipulation by Yingxi Lu, Kairan Dou, Bu He and Mengdi Xu. The project code provides collection and labeling components.

Split	Tasks	Success / failure per task	Trajectories
train	12	200 / 200	4,800
test_ind	the same 12	50 / 50	1,200
test_ood	4 held-out task/variant IDs	50 / 50	400
Total	12 training tasks + 4 OOD IDs		6,400

The older eight-variant, 800-row OOD collection is not included. Counts and prompts are in task_registry.json and dataset_summary.json.

Two interfaces

HF interface: Parquet contains typed Video values with embedded video bytes. No private filesystem path or custom remote dataset script is required.

from datasets import load_dataset
data = load_dataset('luyingxi35/RoboFAPE', revision='v1.0.0')
sample = data['test_ood'][0]
frame = sample['video'].get_frame_at(sample['progress_frame_indices'][0])
progress = sample['target_progress'][0]

Install the recorded compatible datasets, PyTorch and TorchCodec versions, and the FFmpeg runtime required by TorchCodec. To inspect metadata without a decoder:

from datasets import Video
metadata = data['train'].cast_column('video', Video(decode=False))

Exact experiment input: repro/processed/ preserves the original uint8 NPZ arrays, 32 sampled frames per item, labels and language vectors. These arrays are authoritative for reproducing the recorded training/evaluation inputs. Video decoding is not a substitute for these cached arrays, particularly for the re-encoded OOD clips.

import sys
from huggingface_hub import snapshot_download
root = snapshot_download('luyingxi35/RoboFAPE', repo_type='dataset',
                         revision='v1.0.0', local_dir='./RoboFAPE-v1')
sys.path.insert(0, root + '/scripts')
from release_reader import load_processed, read_frames
row = load_processed(root, 'train')[0]
rgb = read_frames(root, row)  # (32, 224, 224, 3), uint8

Use load_native(root) from the same helper to test the Parquet interface offline. Verify the download with python scripts/verify_package.py --root /path/to/RoboFAPE-v1.

Label and frame contract

• target_progress[k] labels frame progress_frame_indices[k] of the published video, not an unspecified original simulator timestep.
• Train/IND contain the frozen released trajectory videos. When the video includes one initial frame before the first label, video_frame_offset=1: label zero refers to video frame one. Other rows have offset zero.
• The four-task OOD videos are already sampled 32-frame clips. Their 32 labels describe these clip frames; this release does not invent missing full-resolution temporal labels or source timestamps.
• Success progress uses the collection's linear 0–1 convention. Failure targets preserve frame-local instant_progress or its validated source representation; they are not replaced by monotonic maximum progress.
• processed_target_progress and the NPZ cache retain the original 32-frame sampling. Sampling can omit the final source frame, so a sampled successful trace need not end at exactly 1.0. partial_success is preserved independently.
• Videos contain rendered simulation observations. The canonical view is render_camera; PutSpoon and PutEggplant originate from 3rd_view_camera. Inspect source_camera for each row.
• This is a video/progress dataset. This release does not provide a uniform raw-action or simulator-state interface.

Field	Meaning
id, task_id, task, split	Original sample identity, exact task prompt and fixed partition
quality_label, partial_success	Preserved outcome labels
video	Embedded video bytes, typed as HF Video
target_progress, progress_frame_indices	Progress and explicit video-frame correspondence
video_num_frames, video_frame_offset, video_representation	Length, initial-frame offset and trajectory-video/clip distinction
processed_target_progress, lang_vector	Frozen cached labels and source language representation; not recomputed embeddings
repro_video, repro_frames, video_sha256	Release-relative reproduction files and video checksum

Layout and reproducibility

data/ contains the three standard splits. repro/train/ and repro/test_ind/ include portable local-HF canonical data and videos; repro/test_ood/ includes the frozen OOD manifest and videos. repro/processed/ contains all exact experiment caches. Media is deliberately available both in Parquet and the reproduction layout.

See TRAINING.md, EVALUATION.md, MODEL_REQUIREMENTS.md, requirements-data.txt, provenance.json and release_report.json. Dataset-only validation does not imply that a new model-training run or a paper-metric replication was performed.

Provenance, limitations and license

The data combine successful simulated demonstrations and failure trajectories with the collection's progress-labeling procedure. These reference labels encode task-specific design choices and conventions; they should not be interpreted as independently error-free measurements. OOD contains four named task/variant IDs and does not establish arbitrary cross-domain coverage. Fixed split labels and the source task prompts are retained.

The training source declares CC BY 4.0; the same declaration is carried into this assembled package subject to the provenance and upstream-material scope recorded in ATTRIBUTION.md. The technical audit is not a certification of third-party asset rights. Code notices are separate from dataset licensing. No model weights or simulator asset bundles are included.

This release is hosted at luyingxi35/RoboFAPE and versioned with the v1.0.0 tag. See CITATION.bib for citation.