SilverKittyy/pi0_so101_libero_finetune

π₀ SO-101 Libero Fine-tuned Model

This is a π₀ (pi-zero) model fine-tuned on the Libero dataset for manipulation tasks with the SO-101 follower arm.

Model Details

• Model Type: π₀ (Physical Intelligence Zero)
• Robot Type: SO-101 Follower Arm
• Dataset: Libero
• Task: Manipulation (pick and place, object manipulation)
• Input: RGB images (224x224) + robot state (32-dim)
• Output: Robot actions (7-dim)

Usage with LeRobot

Recording with Policy

lerobot-record \
  --robot.type=so101_follower \
  --robot.port=/dev/ttyACM0 \
  --robot.cameras='{"up":{"type":"opencv","index_or_path":0,"width":640,"height":480,"fps":30}}' \
  --robot.id=my_awesome_follower_arm \
  --dataset.single_task="pick up the screwdriver" \
  --policy.path=SilverKittyy/pi0_so101_libero_finetune \
  --dataset.repo_id=local_eval \
  --dataset.num_episodes=1

Evaluation

lerobot-eval \
  --policy.path=SilverKittyy/pi0_so101_libero_finetune \
  --dataset.repo_id=your_evaluation_dataset \
  --eval.n_episodes=10

Model Architecture

• Vision Encoder: Frozen vision encoder
• State Processing: 32-dimensional state vector
• Action Space: 7-dimensional (6 joint positions + gripper)
• Normalization: Mean-std normalization for all inputs/outputs

Training Details

• Base Model: π₀ pre-trained model
• Fine-tuning Dataset: Libero manipulation tasks
• Training Steps: 29,999
• Learning Rate: 2.5e-05
• Optimizer: AdamW with cosine decay

Hardware Requirements

• Robot: SO-101 Follower Arm
• Camera: OpenCV-compatible camera (e.g., /dev/video0)
• Compute: CUDA-capable GPU recommended

Safety Notes

• This model is trained for manipulation tasks
• Always ensure proper safety measures when using with real hardware
• Test in simulation or with safety constraints before real-world deployment