Datasets: XDOF / ABC-130k like 55 Follow XDOF 34

You need to agree to share your contact information to access this dataset

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Log in or Sign Up to review the conditions and access this dataset content.

ABC-130k

ABC-130k is the largest open-source robot teleoperation dataset. It contains bimanual manipulation trajectories collected on two-arm YAM stations. Episodes are distributed as MCAP files, with subtask annotations kept as separate artifacts so they can be revised or extended independently of the underlying episode data. For details on the accompanying paper, see abc.bot. Please see the GitHub repo here for code to train and deploy with this dataset.

Dataset Statistics

| Attribute | Value | | ------------------------- | ---------------------------------------------------------- | | Robot | Bimanual station, 2x 6-DoF YAM arms, parallel-jaw grippers | | Cameras | Top camera + 2 wrist cameras (RealSense or ZED-X stations) | | Episode format | MCAP (episode.mcap \+ optional annotation.mcap) | | Video codec | H.264 / H.265 (per-stream; read the format field) | | Resolution | 640x480 (RealSense) / 1920x1200 (ZED-X) | | yam — total hours | 3,598.2 hours | | yam — trajectories | 130,919 episodes | | yam — annotated share | 43,090 episodes |

Dataset Structure

xdof/ABC-130k/
├── README.md
├── LICENSE
├── data/
│   ├── train/                        # training split
│   │   └── task=<task_name>/
│   │       ├── episode_XXXX/
│   │       │   ├── episode.mcap       # trajectory: joint state, gripper, video, calibration, task name
│   │       │   └── annotation.mcap    # subtask labels — annotated episodes only
│   │       └── ...
│   └── val/                          # validation split
│       └── task=<task_name>/
│           └── ...
└── meta/
    ├── tasks.json                     # task list, trajectory counts, hours
    └── stats.json

Each episode is a directory containing one or two MCAP files that share a single absolute time base. An `episode.mcap` file will always be present, and for annotated episodes there will be an additional `annotation.mcap`.

Episode Format

Full channel-level documentation lives in docs/YAM\_DATA\_FORMAT.md. `episode.mcap` carries per-arm joint state and actions (`RobotState`, 6 joints in radians plus a flattened 4x4 end-effector pose), gripper aperture (`GripperState`, 0 = closed, 1 = open), one compressed-video stream per camera, and camera calibration on sibling `…-info` topics. `annotation.mcap` carries the task name on `/instruction`, subtask segment labels on `/subtask-annotation`. The station kinematic/collision model is provided as a MuJoCo MJCF.

Two camera configurations. RealSense stations use a mono top camera (all streams H.264, 640x480). ZED-X stations use a stereo top camera (H.265 top + H.264 wrists, 1920x1200).

Loading the Dataset

Download episodes locally with `huggingface_hub`:

from huggingface_hub import snapshot_download
local_dir = snapshot_download(
    repo_id="xdof/ABC-130k",
    repo_type="dataset",
    allow_patterns=["data/", "meta/", "models/**"],
)

Read an episode with the `mcap` libraries:

pip install mcap mcap-protobuf-support

import numpy as np
from mcap.reader import make_reader
from mcap_protobuf.decoder import DecoderFactory
with open("episode_XXXX/episode.mcap", "rb") as f:
    reader = make_reader(f, decoder_factories=[DecoderFactory()])
    for schema, channel, message, decoded in reader.iter_decoded_messages():
        if channel.topic in ("/left-arm-state", "/right-arm-state"):
            side = "left" if "left" in channel.topic else "right"
            joints = list(decoded.position)               # 6 joint angles (radians)
            ee_pose = np.array(decoded.pose).reshape(4, 4) # world-frame 4x4

Gripper aperture is in `/left-ee-state` and `/right-ee-state`. Streams are sampled on independent clocks — use nearest-neighbor matching on timestamps to associate a video frame with its joint state rather than assuming index alignment. See the format spec for examples.

Timestamps

All timestamps are absolute nanoseconds since the Unix epoch, and `episode.mcap`and `annotation.mcap` share the same base so they align on one timeline. For seconds-from-start, subtract the episode start time (from the `session-metadata`record) or the first message's `log_time`.

License

Released under the Apache License 2.0. See `LICENSE`.

Citation

@article{abc2026,
  title   = {Scalable Behavior Cloning with Open Data, Training, and Evaluation},
  author  = {Allshire, Arthur and Singh, Himanshu Gaurav and Singh, Ritvik and Rashid, Adam and Choi, Hongsuk and McAllister, David and Yu, Justin and Chen, Yiyuan and Huang, Huang and Abbeel, Pieter and Chen, Xi and Duan, Rocky and Isola, Phillip and Malik, Jitendra and Shentu, Fred and Shi, Guanya and Wu, Philipp and Kanazawa, Angjoo},
  year    = {2026},
  journal = {arXiv preprint},
  url = {https://abc.bot/},
}

Copy to bucket new

Downloads last month 281,624

Total file size:

22.1 TB