barx-no-reps-xp3k-pnp-pretrain

BARX No Reps XP-3K PnP Counter to Sink / PnP Sink to Counter Pretrain

Full-parameter source-pretraining checkpoints released with BARX: Cross-Embodiment Transfer via Behavior-Aligned Representations.

Model details

• Method: No Reps
• Representations: action only
• Paper dataset: XP-3K
• Task: PnP Counter to Sink / PnP Sink to Counter
• Source embodiments: IIWA, Kinova3, UR5e
• Training episodes: 6000
• Base VLM: ajaysri/prism-qwen25-extra-dinosiglip-224px-0_5b-stage-finetune-x7
• Training data: ajaysri/robocasa-x-xp3k-pnp
• Training: full-parameter FSDP finetuning

The repository contains the released checkpoint series, resolved configuration, and dataset statistics. Use the BARX release scripts rather than loading the checkpoint as a generic text-generation model: action tokenization, normalization, image preprocessing, and representation prompting are part of the public policy code.

Provenance

The BARX MiniVLA policy combines Qwen2.5-0.5B with DINOv2 and SigLIP vision features through MiniVLA/OpenVLA-derived policy code. This checkpoint was trained in RoboCasa-X using the CC BY 4.0 dataset linked above. Source embodiments are IIWA, Kinova3, and UR5e; Panda, Panda-OG, and Jaco are target embodiments evaluated after adaptation.

Exact training aliases, selected paper checkpoints, adaptation commands, and evaluation conditions are documented in the BARX release repository.

Intended use and limitations

This artifact is intended for simulation robotics research, BARX reproduction, and cross-embodiment adaptation. It has not been validated for real-world robot control or safety-critical use. Short public smoke tests establish executability, not the paper's full success rates or guarantees on unseen robots, scenes, or objects.

License and attribution

This BARX model artifact is released under Apache 2.0. Preserve attribution for Qwen2.5, DINOv2, SigLIP, MiniVLA/OpenVLA, RoboCasa/RoboCasa-X, the training dataset, and BARX. Separately applicable upstream licenses are not replaced by this artifact license.

Links

• Project page
• Paper
• Code, training, and evaluation walkthrough

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}
}