Wjjjh/pi05_PVIA

pi0.5 + UEV Franka deployment bundle

This bundle keeps the official pi0.5 PyTorch model and the UEV visual prompt separate. VPT tokens cannot be folded into the base attention weights.

Files:

• model.safetensors: full pi0.5 model, converted to BF16 from official Orbax weights.
• uev_prompt.pt: four visual prompt tokens inserted at SigLIP layer zero.
• assets/: normalization assets distributed with the official base checkpoint.
• deployment.json: loader metadata.

Load with the modified UEV/OpenPI checkout and config pi05_aloha, passing uev_checkpoint="uev_prompt.pt", uev_num_tokens=4, and uev_layer_index=0 to create_trained_policy.

The supplied normalization assets are official assets, not statistics computed from this dual-arm dataset. Real-robot deployment still requires a matching Franka input/output transform and normalization statistics before commanding hardware.