Hy-Embodied-0.5-VLA

Competition-oriented LingBot-VLA-V2 checkpoint for the Galaxea R1 Pro and the 26D canonical action protocol used by the 2026 Embodied Extreme Challenge.

Training

• Base: robbyant/lingbot-vla-v2-6b at revision 11c703bf6a5c1f45b3b69168482da11fdbba53d7.
• Data: 200 episodes and 32,237 aligned frames from OpenGalaxea/R1Pro_Pick_and_Place_Cube.
• Inputs: three RGB views, a 40D canonical proprioceptive state, an instruction, and an embodiment token.
• Output: 50 steps of canonical 26D delta actions.
• Optimization: one epoch, batch size 2, bf16, FSDP, expert-only fine-tuning; final optimizer step 6,259.

model.safetensors is a full-state checkpoint. Its SHA-256 is 0dc1a8477bf1a8cf59807512eb7e0d2a02651825df26a87eae412790232347b4.

Loading contract

This artifact is not a standalone Transformers model. Use the implementation in FTP2026/embodied, load the upstream LingBot-VLA-V2 release first, then overlay this checkpoint with vla_policy.checkpoints.load_checkpoint. config.json, stats.json, and train_args.json preserve the exact action, normalization, and training contract.

Evaluation status

The checkpoint has passed strict state-dict loading and a fresh-process legal 26D forward test. It has not yet demonstrated task success in the official Isaac Sim closed loop. Treat it as a baseline, not as a safety-qualified robot controller.

License and attribution

The upstream LingBot-VLA-V2 weights are Apache-2.0. The fine-tuning dataset is CC BY-NC-SA 4.0; this derived checkpoint is released under CC BY-NC-SA 4.0. See OpenGalaxea/R1Pro_Pick_and_Place_Cube for source attribution and terms.