jt-2026/pi05-so101-lora-v1

π₀.₅ LoRA Fine-tuned on SO-101

A π₀.₅ (Physical Intelligence) VLA model fine-tuned with dual-expert LoRA on a real SO-101 manipulator arm for breakfast table-setting tasks.

Model info

Field	Value
Base model	π₀.₅ (pi05_base)
Architecture	PaliGemma 3B VLM + 300M action expert
Fine-tuning	LoRA on both VLM and action expert
Training steps	15,000
Dataset	jt-2026/so101-breakfast (49 episodes)
Robot	SO-101 (6-DoF follower arm)
Action space	6-DoF joint positions
Framework	JAX (openpi)

Task

Breakfast table-setting — multi-step, long-horizon manipulation:

"First put the block onto the plate, move the plate to the center of the table, place the spoon on the right side of the plate, and place the cup on the left side of the plate."

Usage with openpi

git clone https://github.com/ljt228/pi05-so101-finetune.git
cd pi05-so101-finetune
git checkout pi05-so101-finetune-0825
uv sync

# Download checkpoint
huggingface-cli download jt-2026/pi05-so101-lora-v1 --local-dir checkpoints/pi05_so101_lora_finetune/so101_lora_v1/15000

# Serve inference
uv run scripts/serve_policy.py --config pi05_so101_lora_finetune --checkpoint checkpoints/pi05_so101_lora_finetune/so101_lora_v1/15000

Training details

• Optimizer: AdamW with EMA disabled
• Batch size: 32
• Learning rate: default openpi schedule
• LoRA rank: default (both VLM and action expert)
• Single NVIDIA L20 GPU

Loss curve

Training converges smoothly — flow-matching loss drops from 0.92 → 0.56 within the first few thousand steps.

Citation

@misc{pi05_so101_2025,
  title={π₀.₅ LoRA Fine-tune on SO-101},
  author={jt-228},
  year={2025},
  howpublished={\url{https://huggingface.co/jt-2026/pi05-so101-lora-v1}}
}