pi0.5-fr3-v3-pick-place

pi0.5 FR3 v3 Pick-and-Place

This checkpoint is a pi0.5-DROID policy fine-tuned for real-world pick-and-place with a Franka Research 3 (FR3) robot. It predicts robot actions from synchronized RGB observations and a natural-language task instruction.

Intended Task

Pick up a cube from a table position and place it into a basket.

Training Data

The model was fine-tuned on the FR3 Pick-and-Place LeRobot Dataset. The demonstrations use a Franka Research 3 robot, two Intel RealSense RGB cameras, and a parallel gripper.

Action Format

This deployment uses the pi0.5-DROID action representation at 15 Hz:

• 7 FR3 joint velocities
• 1 gripper-closedness value

The model predicts the arm trajectory and the timing of gripper close/open transitions. The robot-side runner performs physical gripper commands only at those predicted transitions and verifies a grasp before transport or a release inside the basket.

Evaluation And Safety

This checkpoint is intended for guarded research evaluation, not unrestricted robot operation. The supported deployment uses a local FR3 reference streamer, workspace and joint guards, policy-server warmup, latency checks, physical grasp verification, and an E-stop within reach.

Evaluation results must be recorded against the exact checkpoint identity. Aggregate success metrics are intentionally not reported here until the final v3_12000 validation ledger is complete and reviewed.

Usage

Serve the checkpoint with OpenPI on a GPU machine, then run the guarded client from the robot-control workstation. The full setup, recording, training, and evaluation workflow is documented in the accompanying FR3 robot-learning code repository.

The model requires the same robot configuration, camera arrangement, calibration, action convention, and safety controls used during data collection and evaluation. Do not use it on a different workcell without commissioning that setup first.