ABC-130k → FiftyOne (Native Multimodal MCAP)

Use FiftyOne to explore ABC-130k, the largest open bimanual robot-teleoperation dataset (134,806 episodes / 195 tasks / 3,553 hours). Each episode renders natively from its .mcap: synchronized multi-camera video with robot arm/end-effector telemetry charted against the video timeline.

• This dataset: Voxel51/ABC-130k — load directly with load_from_hub
• Loader (FiftyOne): github.com/Burhan-Q/fiftyone-abc130k
• Full dataset: XDOF/ABC-130k
• Original project: abc.bot · paper · training code

This is an unofficial, modified subset of ABC-130k repackaged for FiftyOne (Apache-2.0, same as the source). The .fo.mcap episodes are not verbatim — see License & attribution. All credit for the dataset belongs to the original authors (see Citation).

Installation

pip install "fiftyone[multimodal]>=1.19.0" huggingface_hub

Native multimodal MCAP support needs fiftyone[multimodal]>=1.19.0 (the multimodal extra and version ≥ 1.19.0), and VFF_MULTIMODAL=1 must be set before fiftyone is imported.

Usage

Want a ready-made validation subset? Load the public Voxel51/ABC-130k dataset. It's a diversity-first subset of the val split (~40 episodes across ~40 tasks, all-H.264, with telemetry plot sidecars) and loads straight into FiftyOne. Load it with:

# NOTE: set VFF_MULTIMODAL=1 environment variable first!

import fiftyone as fo
import fiftyone.utils.huggingface as fouh

dataset = fouh.load_from_hub("Voxel51/ABC-130k", name="ABC-130k", persistent=True)
fo.launch_app(dataset)

Loading the full ABC-130k dataset

Use the loader repo github.com/Burhan-Q/fiftyone-abc130k (abc130k.py + abc130k_reencode.py) to enumerate, download, and ingest any split or custom subset from the gated XDOF/ABC-130k source into a FiftyOne dataset. Follow the same setup instructions outlined above, plus ensure to install ffmpeg (with libx264) and make it available on your PATH.

Command line:

# Diversity-first val subset: one small (H.264) episode from each of 40 tasks (~4 GB)
python abc130k.py --name ABC-130k --split val --max-tasks 40 --episodes-per-task 1

# The entire val split, H.264 only (large; H.265 episodes are skipped)
python abc130k.py --name ABC-130k-full --split val

# Preview the selection without downloading
python abc130k.py --name ABC-130k --split val --max-tasks 40 --dry-run

Flags (--name is required; all others are optional):

Flag	Value	Description
--name	string (required)	FiftyOne dataset name to create or populate
--split	val | train (default val)	Which split to load
--max-tasks	int	Keep at most N distinct tasks, spread across the split for diversity
--episodes-per-task	int	Keep at most K episodes per task
--max-size-mb	float	Skip episodes larger than this before downloading (cheap H.265 pre-filter)
--budget-gb	float	Stop selecting once the cumulative download would exceed this cap
--largest-first	flag (no value)	Prefer larger episodes per task (default prefers smaller / H.264)
--fps-cap	float (default 45)	Re-encode episodes whose camera fps exceeds this to 30 fps; 0 disables
--no-plot	flag (no value)	Skip the JSON <topic>.plot telemetry sidecar channels
--include-h265	flag (no value)	Also load H.265 episodes (they will not render)
--overwrite	flag (no value)	Delete an existing dataset of the same name first
--not-persistent	flag (no value)	Do not persist the dataset across sessions
--data-dir	path (default _data/abc-130k)	Local directory to download episodes into
--token	string	HF token (falls back to the HF_TOKEN env var)
--dry-run	flag (no value)	Print the selection without downloading anything

Python:

import abc130k  # sets VFF_MULTIMODAL before importing fiftyone
import fiftyone as fo

dataset = abc130k.download_and_load("ABC-130k", split="val", max_tasks=40, episodes_per_task=1)
session = fo.launch_app(dataset)

Sensor-data plotting

The loader points each sample at an auto-generated episode.fo.mcap (the raw episode plus JSON <topic>.plot telemetry channels) — without it, the arm/EE telemetry plots nothing. It ≈ doubles on-disk size but is self-contained, so the original episode.mcap files can be deleted to reclaim space.

Every numeric telemetry field (position, velocity, torque, gripper opening, end-effector pose) is stored as a repeated double array, which FiftyOne's scalar-only MCAP Plot panel skips. The loader mirrors each into flat scalar <topic>.plot JSON channels at the source timestamps (on by default; --no-plot disables):

• array fields → per-index scalars: position_0…5, velocity_0…6, torque_0…6, gripper position_0
• 4×4 end-effector pose → pose_x/y/z (translation, m) + pose_roll/pitch/yaw (orientation, rad)

To view: open a sample modal → Add tile → Plot → expand e.g. /left-arm-state.plot → check pose_x/y/z (or joint position_*). The series track the playhead alongside the camera tiles.

Rendering notes (H.264 & frame rate)

• The viewer decodes H.264 only. In ABC-130k, RealSense stations are all-H.264 (render); ZED-X stations are all-H.265 (show "No preview frames"). h264_only=True (default) downloads, detects, and skips H.265 episodes; add --max-size-mb to avoid downloading them, or --include-h265 to keep them.
• High-fps episodes are auto-converted to 30 fps. ~60 fps episodes stall the viewer ("Waiting for H.264 keyframe"); those above --fps-cap (default 45) are re-encoded to 30 fps (GOP-30, no B-frames). source_fps / reencoded_to_30fps record what happened.

Dataset facts

• val split: 189 tasks / 1,671 episodes / ~309 GB; layout data/<split>/<task>/episode_<uuid>/episode.mcap (one .mcap = one sample).
• The public HF tree API returns file paths + LFS sizes without a token, so a size-bounded subset can be planned before authenticating (how --dry-run works).

What you get

One FiftyOne sample per episode, media_type = "multimodal" — Foxglove CompressedVideo cameras (1 top + 2 wrist) plus RobotState/GripperState protobuf telemetry per arm, for both -state and -action. No depth / LiDAR / point clouds. Per-sample fields:

Field	Type	Description
task / episode_id	StringField	Task name / source episode UUID
split	StringField	val or train (also a sample tag)
size_mb	FloatField	Source episode size
station	StringField	mono (RealSense/H.264) or stereo (ZED-X/H.265)
topics / camera_topics	ListField	All MCAP topics / camera streams
plot_topics	ListField	The JSON <topic>.plot telemetry channels
camera_codecs	DictField	{topic: format} (e.g. h264)
duration_s / n_messages	FloatField / IntField	Episode duration / message count
source_fps / reencoded_to_30fps	FloatField / BooleanField	Original fps / whether re-encoded to 30 fps

License & attribution

ABC-130k is released under the Apache License 2.0 by its original authors (abc.bot · amazon-far/abc). This redistribution keeps the same license; see the source Hugging Face dataset page for the original terms.

Modified subset. The .fo.mcap files here are not verbatim ABC-130k episodes: camera streams in the ~60 fps episodes (2 of 40) are re-encoded to 30 fps, and every episode carries added JSON <topic>.plot telemetry channels for the FiftyOne App's Plot panel. All other robot/camera data matches the source. The loader code that produced them lives at github.com/Burhan-Q/fiftyone-abc130k.

Citation

Please use the canonical citation from abc.bot / the paper:

@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},
}