ajaysri/robocasa-x-sp900-panda-og-pnp

RoboCasa-X SP-900 Panda-OG PnP Counter to Sink / PnP Sink to Counter

Processed RLDS/TFDS data released with BARX: Cross-Embodiment Transfer via Behavior-Aligned Representations.

Dataset summary

• —Paper dataset: SP-900
• —Task: PnP Counter to Sink / PnP Sink to Counter
• —Embodiment(s): Panda-OG
• —Episodes: 1800
• —Origin: MimicGen-generated simulator demonstrations
• —TFDS builder id: robocasa_x_sp900_panda_og_pnp
• —Language: English task instructions

Each step exposes a 7-D policy action [dx, dy, dz, droll, dpitch, dyaw, gripper], RGB observations, robot state, 2-D end-effector points, object bounding boxes/names, and BARX language-motion annotations. The terminal action is a deliberate safe-stop value. RGB images are JPEG encoded by TFDS and are visually, but not byte-for-byte, equivalent to the raw HDF5 frames.

Provenance

This repository was converted with the public BARX manifest-driven converter from the checksummed ajaysri/barx-raw-hdf5 archive. The immutable file selection is recorded in the BARX `dataset/subsets/` manifests. Source HDF5 actions use the canonical 12-D simulator layout; the RLDS policy boundary intentionally retains the first seven dimensions.

Generation code and conversion commands are available in the BARX release repository. MimicGen software used for these generated demonstrations has separate NVIDIA non-commercial research/evaluation terms; this released dataset is distributed under the license below.

Intended use and limitations

The dataset is intended for simulation robotics research, BARX reproduction, representation learning, format conversion, and cross-embodiment transfer. It is not a real-robot safety dataset, does not establish real-world performance, and should not be used as the sole basis for deploying a physical robot.

License and attribution

This dataset is released under CC BY 4.0. Preserve this card's provenance and cite the BARX paper when redistributing the data or a derived format. Upstream RoboCasa assets and software retain their own terms.

Links

• —Project page
• —Paper
• —Code and release walkthrough
• —Raw HDF5 source

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}
}