rlinf-pi05-panda-joint8-exp02

RLinf Pi0.5 Panda Joint8 EXP-02

This repository is the versioned model bundle used by XPolicyLab's policy/RLinf_Pi05 adapter.

Architecture

• Pi0.5 flow-matching VLA
• 32-D model state/action space
• SigLIP So400m/14 visual encoder
• PaliGemma 2B language expert
• Gemma 300M action expert
• 10-step action horizon
• 200-token prompt/state sequence

Robot-specific vectors occupy the leading dimensions; unused dimensions are zero-padded. This checkpoint uses eight physical dimensions: seven Panda arm joint deltas and one absolute binary gripper channel.

Files

• model.safetensors: bfloat16 standalone Pi0.5 state dict
• norm_stats.json: 32-D mean/std normalization statistics
• paligemma_tokenizer.model: PaliGemma sentencepiece tokenizer
• manifest.json: architecture, action semantics, source-checkpoint digest, file sizes, and SHA256 checksums

Source Checkpoint

The exported source was RLinf global_step_2000/full_weights.pt with SHA256:

fc7f040af64c6db842694b37e3548fb998d6595a0f0c923abb8ff06044061a49

The bfloat16 export SHA256 is recorded in manifest.json.

Usage

Use the adapter and Docker runtime from:

https://github.com/soul667/XPolicyLab/tree/main/policy/RLinf_Pi05

The adapter downloads this repository automatically. To pre-download it:

hf download luobai/rlinf-pi05-panda-joint8-exp02 \
  --local-dir /path/to/artifacts

Then set:

export RLINF_PI05_ARTIFACT_DIR=/path/to/artifacts

Action Contract

• Model space: 32-D
• Physical space for this bundle: 8-D
• Arm dimensions: first 7 values are relative joint deltas
• Gripper dimension: last value is absolute 0=open, 1=closed
• Normalization: mean/std

The XPolicyLab adapter adds current arm joints after unnormalization and returns standard XPolicyLab action dictionaries.