sii-rhos-ai/vifailback-dataset-lerobot

ViFailback Dataset — LeRobot

This repository is a LeRobot v2.1 conversion of the trajectory portion of sii-rhos-ai/ViFailback-Dataset, introduced in the CVPR 2026 paper Diagnose, Correct, and Learn from Manipulation Failures via Visual Symbols.

ViFailback contains real-world ALOHA dual-arm manipulation trajectories designed for studying failure diagnosis, failure localization, corrective guidance, recovery, and learning from failed demonstrations. This conversion keeps the robot trajectories in the standard LeRobot layout and preserves the ViFailback annotations in both frame-level features and episode-level metadata.

Dataset Summary

Statistic	Value
Episodes	5,202
Frames	1,297,146
Tasks	100
Successful episodes	657
Failed episodes	4,545
Avoid keyframes	4,545
Correct keyframes	4,545
RGB cameras	3
Videos	15,606
Frame rate	25 FPS
Approximate trajectory duration	14.41 hours
Local converted size	approximately 28 GiB

Each failed episode has one annotated avoid keyframe and one annotated correct keyframe. The failure distribution is:

Episode outcome / failure type	Episodes	Percentage of all episodes
Success	657	12.63%
Task Planning	645	12.40%
Gripper 6D-Pose	2,771	53.27%
Gripper State	988	18.99%
Human Intervention	141	2.71%

What Is Included

This repository contains:

• all 5,202 ALOHA trajectories as episode-level Parquet files;
• three synchronized RGB streams per episode, encoded as AV1 MP4 video;
• joint state, joint velocity, joint effort, target joint action, end-effector action, and base action;
• the English task instruction associated with every frame through task_index and meta/tasks.jsonl;
• frame-level failure and keyframe flags;
• the complete source episode annotations in meta/vifailback_annotations.jsonl.

Depth maps, action_leader, standalone annotation images, and the ViFailback-Bench/VQA samples are not represented as LeRobot features in this conversion. Fields in the preserved annotations that point to source images or videos are provenance paths relative to the original ViFailback repository; those files are not duplicated here.

Dataset Structure

vifailback-dataset-lerobot/
├── data/
│   └── chunk-{episode_chunk:03d}/
│       └── episode_{episode_index:06d}.parquet
├── videos/
│   └── chunk-{episode_chunk:03d}/
│       ├── observation.images.cam_high/
│       │   └── episode_{episode_index:06d}.mp4
│       ├── observation.images.cam_left_wrist/
│       │   └── episode_{episode_index:06d}.mp4
│       └── observation.images.cam_right_wrist/
│           └── episode_{episode_index:06d}.mp4
└── meta/
    ├── info.json
    ├── tasks.jsonl
    ├── episodes.jsonl
    ├── episodes_stats.jsonl
    └── vifailback_annotations.jsonl

The dataset uses six chunks with up to 1,000 episodes per chunk and a single train split covering episodes 0:5202. No evaluation split is imposed so users can construct task-level, trajectory-level, or failure-type-specific splits appropriate for their experiments.

LeRobot Features

Feature	Type and shape	Description
observation.state	float32[14]	Current ALOHA puppet-arm joint positions (qpos).
action	float32[14]	Target joint positions; the source dataset defines this as qpos at the next step.
observation.velocity	float32[14]	Joint velocities (qvel).
observation.effort	float32[14]	Joint effort values.
complementary_info.action_eef	float32[16]	Target end-effector poses and gripper states for the two arms.
complementary_info.base_action	float32[2]	Robot base commands [base_x, base_y].
complementary_info.vifailback_flags	bool[3]	Episode failure and annotated keyframe flags described below.
observation.images.cam_high	video [3, 480, 640]	High-view RGB camera, AV1/yuv420p at 25 FPS.
observation.images.cam_left_wrist	video [3, 480, 640]	Left wrist RGB camera, AV1/yuv420p at 25 FPS.
observation.images.cam_right_wrist	video [3, 480, 640]	Right wrist RGB camera, AV1/yuv420p at 25 FPS.
timestamp	float32[1]	Time within the episode in seconds.
frame_index	int64[1]	Zero-based frame index within the episode.
episode_index	int64[1]	Zero-based episode index.
index	int64[1]	Global frame index.
task_index	int64[1]	Index into meta/tasks.jsonl.

For each 14-dimensional joint vector, indices 0:7 describe the left arm and indices 7:14 describe the right arm. Each arm uses six joints followed by one gripper value. The 16-dimensional end-effector vector contains [x, y, z, qx, qy, qz, qw, gripper] for the left arm followed by the same values for the right arm.

Frame-Level ViFailback Flags

complementary_info.vifailback_flags is ordered as:

[is_failure_episode, is_avoid_keyframe, is_correct_keyframe]

• is_failure_episode is true for every frame of a failed episode and false for every frame of a successful episode.
• is_avoid_keyframe is true only at the source annotation's avoid keyframe.
• is_correct_keyframe is true only at the source annotation's correct keyframe.

This representation supports direct frame-level sampling without parsing JSON metadata.

ViFailback Episode Metadata

meta/vifailback_annotations.jsonl contains one JSON object per episode, ordered by episode_index. Every object has the following top-level fields:

Field	Description
episode_index	Episode index in this LeRobot dataset.
source_hdf5	Relative path of the original HDF5 trajectory.
source_task	Original task-directory name.
source_episode	Original episode name.
source_repo_id	sii-rhos-ai/ViFailback-Dataset.
num_frames	Number of frames in the episode.
episode_success	Normalized value: success or failure.
avoid_keyframe_indices	Normalized zero-based avoid-keyframe indices.
correct_keyframe_indices	Normalized zero-based correct-keyframe indices.
annotation	Complete original annotation object.

The nested annotation object preserves the original fields task, failure_detection, subtasks, failure_subtask, failure_type, keyframe, avoidance, correction, images, video, and source_index. In particular, the avoidance and correction entries retain the low-level natural-language guidance and visual-symbol annotation code.

Loading the Dataset

The metadata reports LeRobot dataset format v2.1. Use a LeRobot release that can read v2.1 datasets.

from lerobot.common.datasets.lerobot_dataset import LeRobotDataset

dataset = LeRobotDataset(
    repo_id="sii-rhos-ai/vifailback-dataset-lerobot",
    download_videos=True,
)

print(dataset)
sample = dataset[0]

state = sample["observation.state"]
action = sample["action"]
cam_high = sample["observation.images.cam_high"]
is_failure, is_avoid, is_correct = sample[
    "complementary_info.vifailback_flags"
]

The episode-level annotations can be read after the repository metadata has been downloaded:

import json

annotation_path = dataset.root / "meta" / "vifailback_annotations.jsonl"
with annotation_path.open(encoding="utf-8") as f:
    annotations = [json.loads(line) for line in f]

episode_0_annotation = annotations[0]
print(episode_0_annotation["episode_success"])
print(episode_0_annotation["annotation"]["failure_type"])

AV1 decoding support is required to read the RGB videos. If video decoding fails, check that the selected LeRobot video backend and the system FFmpeg installation support AV1.

Conversion Notes

• Episode ordering is deterministic: task directories are sorted lexicographically and episode numbers are sorted numerically within each task.
• The complete annotations are retained without rewriting their nested contents; normalized outcome and keyframe fields are added at the top level for convenient access.
• RGB frames are decoded without an additional BGR/RGB channel swap, following the storage note in the original ViFailback dataset card.
• Depth data was intentionally omitted from this conversion. The original dataset notes that RGB images and depth maps from Dabai cameras are not spatially aligned and require preprocessing before RGB-D fusion or point-cloud generation.
• Video encoding is lossy AV1. Users requiring the original compressed frames or raw depth should use the source dataset.

Intended Uses

This dataset is intended for research on:

• robot imitation learning and task-conditioned manipulation;
• manipulation failure detection and classification;
• temporal failure and keyframe localization;
• failure-aware sampling and representation learning;
• corrective-guidance and recovery-policy learning;
• multimodal reasoning over robot state, action, language, and video.

Limitations and Responsible Use

• The data comes from a particular ALOHA setup, camera configuration, workspace, object distribution, and collection procedure. Results may not transfer directly to other robots or environments.
• Failed trajectories substantially outnumber successful trajectories, and Gripper 6D-Pose failures are the dominant category. Account for this imbalance during training and evaluation.
• is_failure_episode is an episode-level label repeated on all frames; it does not mean that every frame visually depicts the failure.
• Avoid and correct keyframes are sparse annotations and should not be interpreted as exhaustive temporal boundaries of a failure event.
• The actions are recorded robot commands/targets, not guarantees of safe execution. Policies trained on this dataset require independent safety validation before deployment on physical hardware.
• This conversion focuses on robot trajectories. Refer to the original ViFailback repository for benchmark-specific and VQA assets.

Source, License, and Citation

This is a format conversion of sii-rhos-ai/ViFailback-Dataset. The source dataset is released under the MIT License. Please follow the source repository's terms and cite the original work when using this conversion.

@inproceedings{zeng2026diagnose,
  title={Diagnose, Correct, and Learn from Manipulation Failures via Visual Symbols},
  author={Zeng, Xianchao and Zhou, Xinyu and Li, Youcheng and Shi, Jiayou and Li, Tianle and Chen, Liangming and Ren, Lei and Li, Yong-Lu},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={42386--42395},
  year={2026}
}