RLinf-Pi05-BEHAVIOR-1K-PT50-CS32

RLinf-Pi05-BEHAVIOR-1K-PT50-CS32

This repository contains the shared Pi0.5 BEHAVIOR checkpoint used by the RPent BEHAVIOR integration.

The RPent policy profile id is:

pi05-b1kpt50-cs32

Files

The runtime-critical files are:

model.safetensors
config.json
assets/behavior-1k/2025-challenge-demos/norm_stats.json

A duplicate normalization statistics file is also included at:

physical-intelligence/behavior/norm_stats.json

This duplicate mirrors the LIBERO-style physical-intelligence/... checkpoint layout. The canonical BEHAVIOR runtime path remains assets/behavior-1k/2025-challenge-demos/norm_stats.json.

Checksums

7e257666d835f6af701de493676a6c86a0421b2efc737a0f911d782b7a09f635  model.safetensors
a4ae208203adfdd64c5fdbd4b0dc257e4ebbc82e464cb146dd0377051b25fc0a  config.json
d66ed16830a98f90dde8a315058b4a0df59f5e05734c1686d8b3f66787d0a929  assets/behavior-1k/2025-challenge-demos/norm_stats.json
d66ed16830a98f90dde8a315058b4a0df59f5e05734c1686d8b3f66787d0a929  physical-intelligence/behavior/norm_stats.json

Configuration

config.json:

{
  "action_dim": 32,
  "action_horizon": 32,
  "paligemma_variant": "gemma_2b",
  "action_expert_variant": "gemma_300m",
  "precision": "bfloat16"
}

RPent BEHAVIOR uses the pi05_behavior OpenPI configuration. The policy emits 32-dimensional chunks and RPent maps them to the BEHAVIOR environment action contract.

Download

Install the Hugging Face CLI:

pip install -U huggingface_hub

Download the checkpoint:

hf download RLinf/RLinf-Pi05-BEHAVIOR-1K-PT50-CS32 \
  --local-dir ./checkpoints/RLinf-Pi05-BEHAVIOR-1K-PT50-CS32

export PI05_CHECKPOINT_PATH="$PWD/checkpoints/RLinf-Pi05-BEHAVIOR-1K-PT50-CS32"

Optional checksum verification:

cd "$PI05_CHECKPOINT_PATH"
sha256sum model.safetensors config.json \
  assets/behavior-1k/2025-challenge-demos/norm_stats.json \
  physical-intelligence/behavior/norm_stats.json

Intended Use

This checkpoint is intended for research use with RPent BEHAVIOR. It is not a general-purpose language model and should be evaluated only through the BEHAVIOR/RPent robotics stack.

For BEHAVIOR evaluation, official task success is determined by the simulator's task success signal, exposed in RPent as task_success from info["done"]["success"]. Primitive-level success, reward, video appearance, or local workflow completion are not substitutes for official success.

Related Links

• RPent: https://github.com/RLinf/RPent
• RLinf: https://github.com/RLinf/RLinf
• RLinf Hugging Face organization: https://huggingface.co/RLinf
• LIBERO counterpart: https://huggingface.co/RLinf/RLinf-Pi05-LIBERO-130-fullshot-SFT