Datasets:

ajaysri
/

barx-raw-hdf5

license: cc-by-4.0
language:
  - en
pretty_name: BARX Raw HDF5 Demonstrations
size_categories:
  - 10K<n<100K
tags:
  - robotics
  - imitation-learning
  - hdf5
  - robocasa
  - mimicgen
  - barx

BARX Raw HDF5 Demonstrations

This repository contains the normalized raw simulator demonstrations released with BARX: Cross-Embodiment Transfer via Behavior-Aligned Representations. It is intended for users who want to inspect the source trajectories, convert them into another training format, reproduce the BARX RLDS conversion, or prepare source demonstrations for new data-generation work.

Dataset summary

23,400 demonstrations in 240 HDF5 files
approximately 284.03 GiB
22,800 MimicGen-generated demonstrations and 600 human-teleoperated simulator demonstrations
six embodiments: IIWA, Jaco, Kinova3, Panda, Panda-OG, and UR5e
four tasks with 5,850 demonstrations each: pick-and-place counter-to-sink, pick-and-place sink-to-counter, turn-on-sink-faucet, and flip-mug-upright
paper subsets: XP-900, XP-3K, SP-900, and target-50

The human subset records simulated robot teleoperation; it contains no camera recordings or personal information from the human operators.

Repository layout

human/                 # 12 target/task files, 50 demonstrations each
mg/                    # 228 generated files, 100 demonstrations each
manifest.csv           # paths, sizes, SHA-256 hashes, metadata, and paper sets

Each file stores demonstrations below data/demo_*. Principal fields include:

actions: [T, 12] canonical simulator actions;
states: simulator states for replay and data generation;
obs/agentview_rgb: [T, 180, 320, 3] RGB observations;
obs/ee_states and obs/gripper_states: policy state inputs;
aux_info/eef_normalized_image_pts and aux_info/bboxes_2d/*: BARX representation annotations; and
ep_meta: JSON episode metadata including the language instruction.

The canonical action layout is [arm(6), gripper(1), base(3), torso(1), mode(1)]. Stored simulator asset paths use the portable <ROBOCASA>/ token rather than installation-specific paths.

Download and verification

Download the full repository with the Hugging Face CLI:

hf download ajaysri/barx-raw-hdf5 --repo-type dataset --local-dir barx-raw-hdf5

The BARX code release provides a manifest-driven verifier and RLDS converter. See the BARX project page for the release code, installation instructions, and pinned dataset revision.

Intended use

The dataset is intended for robotics and imitation-learning research, cross-embodiment representation learning, format conversion, simulator replay, and preparation of new data-generation pipelines. Users should preserve the included attribution and cite BARX and RoboCasa in resulting work.

Limitations

These are simulator demonstrations and do not directly characterize real-world robot performance.
The task and embodiment distribution is intentionally narrow and should not be treated as a general-purpose robotics corpus.
The raw demonstrations preserve the information needed to prepare MimicGen source datasets, but the separately pinned BARX MimicGen generator and task configurations are required for end-to-end regeneration.
HDF5 is provided for exact source preservation and flexible conversion; the processed BARX RLDS repositories are more convenient for policy training.

License and attribution

The BARX dataset is released under the Creative Commons Attribution 4.0 International license. Rendered simulator observations incorporate RoboCasa environments and assets; see RoboCasa for its attribution and upstream terms.

Citation

@inproceedings{sridhar2026barx,
  title     = {Cross-Embodiment Transfer via Behavior-Aligned Representations},
  author    = {Sridhar, Ajay and Gao, Jensen and Yang, Jonathan and Mercat, Jean and Belkhale, Suneel and Sadigh, Dorsa},
  booktitle = {IEEE International Conference on Robotics and Automation (ICRA)},
  year      = {2026}
}

Please also cite RoboCasa and MimicGen when using the corresponding simulator assets or generated subset.