README.md

---
license: apache-2.0
pretty_name: ABC
language:
  - en
tags:
  - robotics
  - manipulation
  - imitation-learning
  - bimanual
  - teleoperation
  - mcap
task_categories:
  - robotics
size_categories:
  - n>1T
configs:
  - config_name: yam
    data_files:
      - split: train
        path: data/train/**
      - split: val
        path: data/val/**
---

# 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](https://abc.bot/).
Please see the [GitHub repo here](https://github.com/amazon-far/abc) for code to
train and deploy with this dataset.

<video src="https://huggingface.co/datasets/xdof/ABC-130k/resolve/main/images/yam_data_example.mp4" controls autoplay loop muted playsinline width="100%"></video>

## 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** | 3,590.7 hours |
| **yam — trajectories** | 130,703 episodes |
| **yam — annotated share** | 42,980 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`](./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`:

```python
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
```

```python
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

```bibtex
@misc{abc2026,
  title         = {Scalable Behavior Cloning with Open Data, Training, and Evaluation},
  author        = {Arthur Allshire and Himanshu Gaurav Singh and Ritvik Singh and Adam Rashid and Hongsuk Choi and David McAllister and Justin Yu and Yiyuan Chen and Huang Huang and Pieter Abbeel and Xi Chen and Rocky Duan and Phillip Isola and Jitendra Malik and Fred Shentu and Guanya Shi and Philipp Wu and Angjoo Kanazawa},
  year          = {2026},
  eprint        = {2606.27375},
  archivePrefix = {arXiv},
  primaryClass  = {cs.RO},
  doi           = {10.48550/arXiv.2606.27375},
  url           = {https://arxiv.org/abs/2606.27375},
}
```