OpenArm Pi0.5 Finetuned Checkpoint
This is a finetuned Pi0.5 model checkpoint for the OpenArm bimanual robot.
Model Details
- Base Model: Pi0.5 (Physical Intelligence)
- Fine-tuned on: OpenArm pick-and-place dataset
- Action Dimension: 16 (2 arms × (7 joints + 1 gripper))
- Action Horizon: 16 timesteps
- Training Framework: OpenPI (JAX/Flax)
Usage
Installation
git clone https://github.com/AiSaurabhPatil/openpi.git
cd openpi
uv sync
Download & Inference
from huggingface_hub import snapshot_download
# Download checkpoint
checkpoint_path = snapshot_download(
repo_id="AiSaurabhPatil/openarm-pi05-finetuned",
local_dir="./checkpoints/pi05_openarm/openarm_pi0.5_finetuned/5000"
)
# Run inference server
# Terminal 1:
# XLA_PYTHON_CLIENT_MEM_FRACTION=0.55 uv run scripts/serve_policy.py policy:checkpoint \
# --policy.config=pi05_openarm \
# --policy.dir=./checkpoints/pi05_openarm/openarm_pi0.5_finetuned/5000
# Terminal 2:
# python scripts/openarm_policy_client.py \
# --task="grab the electric screwdriver from the table and drop it inside the box"
Isaac Sim Integration
This checkpoint is designed to work with Isaac Sim via ROS2:
- Run the Isaac Sim OpenArm bridge
- Start the policy server with this checkpoint
- Run the policy client for inference
Training Details
- Dataset: Custom OpenArm demonstrations recorded via teleoperation
- Training Steps: 5,000+
- Batch Size: 32
- Learning Rate: 2.5e-5 (cosine decay)
- Hardware: NVIDIA A100 80GB
Files
├── params/ # Model weights (Orbax checkpoint format)
├── assets/ # Normalization statistics
└── _CHECKPOINT_METADATA # Checkpoint metadata
Citation
If you use this checkpoint, please cite:
@software{openpi2025,
author = {Physical Intelligence},
title = {OpenPI: Open-source Pi Foundation Models},
year = {2025},
url = {https://github.com/Physical-Intelligence/openpi}
}
License
Apache 2.0