yangzhixing/pi0fast-base-lora-so101-box-to-plate

PI0-FAST LoRA for SO-101 Box-to-Plate

This repository contains a LoRA adapter fine-tuned from lerobot/pi0fast-base on maxlium/so101-box-to-plate.

The policy consumes SO-101 front and wrist camera observations, a 6-dimensional robot state, and a task instruction. It predicts 10-step chunks of 6-dimensional SO-101 actions.

Training

• Method: LoRA on PI0-FAST language-model attention q_proj and v_proj
• LoRA rank / alpha: 16 / 16
• Trainable parameters: 1,843,200
• Precision: bfloat16
• Selected checkpoint: 12,000 steps
• Split: 80% training episodes / 20% validation episodes
• Validation loss at 12,000 steps: 2.6770

The 12,000-step checkpoint had the lowest validation loss among saved checkpoints. This loss is an offline action-token cross-entropy metric and does not replace real-robot success evaluation.

Usage

Install LeRobot with PEFT support, then pass this repository as the policy path:

pip install "lerobot[peft]"

lerobot-record \
  --policy.path=yangzhixing/pi0fast-base-lora-so101-box-to-plate \
  <your SO-101 robot and camera arguments>

The adapter automatically loads the lerobot/pi0fast-base base model. It expects observations named observation.images.front, observation.images.wrist, and observation.state.

Limitations and safety

This adapter is specialized for the camera setup, action representation, and task distribution in the training dataset. Validate camera keys, calibration, and action bounds before deployment. Begin real-robot testing at reduced speed with an emergency stop available.