pi05-droid-franka-lora

Fine-tuned pi0.5 (openpi) checkpoint for the TASL Franka FR3 bench. LoRA fine-tuned on 16 GELLO-teleoperated episodes (4,262 frames, 15 Hz) of "pick blue cup and put into red cup" pick-and-place on a Franka Research 3 with Robotiq 2F-85 gripper and dual ZED cameras (exterior ZED 2i + wrist ZED Mini).

Base model: pi05_droid (Physical Intelligence, DROID-pretrained). Training used openpi's pi05_droid_franka_lora config: 20,000 steps (checkpoint saved at step 19,999), global batch 32, LoRA on the PaliGemma 2B backbone and the 300M action expert. Norm-stats are the original DROID stats (asset_id=droid) and must be reused at inference.

Layout

• params/ — orbax (JAX) model weights
• assets/ — norm stats (droid/norm_stats.json)
• train_state/ — optimizer state (not needed for inference)
• _CHECKPOINT_METADATA — orbax checkpoint metadata

Serve with openpi

python scripts/serve_policy.py --port 8000 \
  policy:checkpoint \
  --policy.config=pi05_droid_franka_lora \
  --policy.dir=<path to this repo after download>

The pi05_droid_franka_lora config must be registered in the openpi checkout you serve from (src/openpi/training/config.py).

Training snapshot

• Checkpoint step: 19,999 / 20,000
• Data: tasl/test_finetune, 16 episodes, 2 task labels, 15 Hz, 2 cameras, 224x224 RGB
• Final loss: ~0.0024 (may indicate overfitting on this small dataset)

Fine-tune of the openpi pi05_droid checkpoint — check upstream openpi and Physical Intelligence license terms before redistribution or commercial use.