ajaysri/robocasa-x-xp900-pnp-vq-extra-action-tokenizer

RoboCasa-X XP-900 PnP VQ Action Tokenizer

VQ action tokenizer used by the public BARX XP-900 pick-and-place walkthrough and source-pretraining checkpoints.

Model details

• CLI tokenizer id: robocasa-x-xp900-pnp-vq-extra-action-tokenizer
• Task: PnP Counter to Sink / PnP Sink to Counter
• Source embodiments: IIWA, Kinova3, UR5e
• Training data: ajaysri/robocasa-x-xp900-pnp
• Action dimensions: 7
• Action chunk: 8 actions (future_action_horizon=7)
• Codebook: 256 codes, 7 residual groups, 512 latent dimensions
• Expected BARX directory: vq/mg_pnp_lite

The tokenizer converts normalized 7-D policy-action chunks into discrete action tokens. It is task/dataset specific and should be used with the normalization statistics and policy configuration supplied by the BARX release.

Provenance

The tokenizer was trained with the public BARX VQ-BeT-derived training path on the CC BY 4.0 XP-900 PnP dataset linked above. That dataset contains IIWA, Kinova3, and UR5e source demonstrations for both pick-and-place directions. Configuration and training commands are recorded in the BARX release repository.

Intended use and limitations

This artifact is intended for BARX simulation-policy training and evaluation. It is not a standalone robot controller, is not validated on real robots, and should not be substituted across tasks or action-normalization regimes without retraining and validation.

License and attribution

This BARX model artifact is released under Apache 2.0. The training data is CC BY 4.0. Preserve attribution for the BARX data, VQ-BeT lineage, RoboCasa-X, MiniVLA/OpenVLA, and BARX.

Links

• Project page
• Paper
• Training and evaluation documentation

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}
}