CRIA: Contact-Rich Interaction Anomaly Anticipation Dataset

🌐 Project page: joringell.github.io/project/MAIA
📄 Paper: arXiv:2610.21686

CRIA is a real-world multimodal dataset for studying anomaly anticipation in contact-rich robot manipulation. It contains synchronized multi-view RGB videos, robot state and action signals, force/torque measurements, interaction-onset annotations, and frame-level anomaly labels.

The dataset contains 2,852 trajectories from 18 manipulation tasks in 5 task categories. Among them, 1,817 trajectories contain at least one annotated anomalous event, with 2,328 anomaly events in total.

Dataset statistics

Category	Trajectories	Trajectories with anomalies	Anomaly events	Anomalous-frame ratio
Insertion	600	540	709	42.11%
Wiping	500	260	286	15.59%
Pressing	600	330	418	21.57%
Path-constrained manipulation	541	385	559	13.28%
Articulated-object manipulation	611	302	356	14.04%
Total	2,852	1,817	2,328	20.88%

Each task contains at least 150 trajectories, including human demonstrations, human demonstrations with diverse anomalous interactions, and autonomous rollouts generated by the π0.5 policy. In the released metadata, the corresponding source values are human and pi05_joint.

Task categories

Category	Tasks
Insertion	insert_ni, insert_plug_2, insert_plug_3, insert_typec
Wiping	wipe_board, wipe_bowl, wipe_vase
Pressing	press_bottle, press_seal, press_switch1, press_switch2
Path-constrained manipulation	pushc, pushi, zipper
Articulated-object manipulation	microwave, open_box, open_box2, pull_drawer

The tasks cover precise insertion and alignment, maintaining contact while wiping surfaces, pressing objects with controlled force, following constrained manipulation paths, and manipulating articulated objects such as drawers, boxes, and a microwave.

Data collection

The demonstrations were collected with a Franka FR3 robot equipped with an ATI Axia80-M8 six-axis force/torque sensor at the wrist. Three Intel RealSense D435i cameras record synchronized RGB observations from the wrist and two external viewpoints. In this release, the video keys are: wrist, side_1, side_2.

RGB videos are stored at 30 Hz and 256 × 192 resolution. Robot state and action signals are aligned to the 30 Hz control stream. Force/torque measurements are recorded at 120 Hz and aligned with the 30 Hz control timeline during preprocessing. The released episode records preserve force/torque measurements in chunked arrays..

Human demonstrations were collected with leader–follower teleoperation and real-time force feedback. The dataset combines successful and failed task outcomes with anomalous interactions, making it suitable for anomaly detection, anomaly anticipation, contact-rich manipulation, and multimodal robot learning.

Annotations

The release contains both episode-level and frame-level annotations.

• episode_label: episode outcome. In this release, 0 denotes failure and 1 denotes success.
• frame_label: frame-level anomaly annotation. 0 denotes an anomalous frame and 1 denotes a nominal frame.
• interaction_label: interaction-state annotation. It identifies the task-relevant interaction phase and its onset. For insertion, pressing, and wiping tasks, onset is defined around the moment the end-effector reaches the target vicinity before contact. For articulated and path-constrained tasks, onset is defined after the task-relevant grasp or contact has been established.

Episode success/failure and anomalous interaction are different concepts: an episode can contain an anomaly without every frame or every task outcome being equivalent to the episode-level label.

Directory structure

CRIA is released as a collection of 18 task-level datasets. Each task preserves the LeRobot v3.0 layout:

CRIA/
├── insert/
│   ├── insert_ni/
│   │   ├── data/
│   │   │   └── chunk-*/episode_*.parquet
│   │   ├── meta/
│   │   │   ├── info.json
│   │   │   ├── episodes.jsonl
│   │   │   ├── episodes_stats.jsonl
│   │   │   ├── annotation.json
│   │   │   ├── stats.json
│   │   │   └── tasks.jsonl
│   │   └── videos/
│   │       ├── wrist/chunk-*/episode_*.mp4
│   │       ├── side_1/chunk-*/episode_*.mp4
│   │       └── side_2/chunk-*/episode_*.mp4
│   └── ...
├── wipe/
├── press/
├── path/
├── articulated/
└── scripts/
    ├── check_dataset.py
    └── visualize.py

The metadata schema follows LeRobot v3.0. Important fields include:

Group	Representative fields
Episode/frame indexing	episode_index, frame_index, index, timestamp, timestamp_us
Outcome and annotations	episode_label, frame_label, interaction_label
Robot state	observation.state.ee_pose, observation.state.gripper_state, observation.state.q
Commands and actions	action, observation.state.ee_pose_cmd, observation.state.q_cmd
Force/torque	observation.state.wrench_raw_chunk, observation.state.wrench_filtered_chunk, observation.state.wrench_chunk_valid_len
RGB observations	wrist, side_1, side_2
Dataset metadata	source_id, task_index

Download

Using the Hugging Face CLI:

hf download Joringell/CRIA \
  --repo-type dataset \
  --local-dir CRIA

Using Python:

from huggingface_hub import snapshot_download

root = snapshot_download(
    repo_id="Joringell/CRIA",
    repo_type="dataset",
    local_dir="CRIA",
)
print(root)

Load a downloaded task locally

After downloading CRIA, load one task from its local LeRobot directory. The root argument must point to a specific task, not to the CRIA repository root:

from lerobot.datasets.lerobot_dataset import LeRobotDataset

dataset = LeRobotDataset(
    repo_id="CRIA/articulated_microwave",
    root="/data/CRIA/articulated/microwave",
)

sample = dataset[0]
print(sample.keys())

Here, repo_id is only a local dataset identifier because the data has already been downloaded. The local files are found through root.

For direct access to an episode Parquet file without using the LeRobot loader:

import pandas as pd

episode = pd.read_parquet(
    "/data/CRIA/articulated/microwave/"
    "data/chunk-000/episode_000000.parquet"
)

print(episode.head())

Dataset validation

Run the included checker after downloading or copying the dataset:

python scripts/check_dataset.py

When running it from another directory, provide the dataset root explicitly:

python /path/to/CRIA/scripts/check_dataset.py --root /path/to/CRIA

The checker verifies task discovery, metadata readability, data/video counts, episode metadata counts, and the absence of backup files. The released dataset currently contains 2,852 parquet episodes, 8,556 MP4 videos, 2,852 episode records, and 2,852 episode-stat records.

Visualization

The repository includes a local browser-based viewer. It displays the three synchronized videos, force/torque signals, end-effector position, interaction onset, and frame-level anomaly regions.

From the CRIA directory:

python scripts/visualize.py

From any other directory:

python /path/to/CRIA/scripts/visualize.py --root /path/to/CRIA

Useful options:

python scripts/visualize.py --help
python scripts/visualize.py --verbose
python scripts/visualize.py --host 0.0.0.0 --port 7860

The viewer supports filtering by task, source, result, and episode. Press Enter to load the selected episode and Space to play or pause it.

Intended uses

CRIA is intended for research on:

• multimodal anomaly detection and anomaly anticipation;
• contact-rich robot manipulation;
• force/torque-aware policy learning;
• visual, proprioceptive, and force/torque signal fusion;
• robot demonstration and policy-rollout analysis.

Limitations

CRIA was collected in controlled laboratory environments using a specific robot platform, sensor configuration, and set of manipulation tasks and objects. Models trained on CRIA may not generalize directly to other robots, environments, objects, or sensor configurations. Users should consider task-specific characteristics and annotation protocols when interpreting results or comparing performance across tasks.

License

CRIA is released under the Creative Commons Attribution 4.0 International License (CC BY 4.0).

Citation

The paper citation and arXiv identifier will be added after the preprint is released. For now, please cite the dataset and project page:

@misc{qiao2026maia,
title={MAIA: Anticipating Interaction Anomalies in Contact-Rich Manipulation via Multimodal World Models},
author={Yi Qiao and Maolin Zheng and Haoran Sun and Lujia Dong and Wenzhao Lian},
year={2026},
eprint={2610.21686},
archivePrefix={arXiv},
primaryClass={cs.RO},
url={https://arxiv.org/abs/2610.21686},
}