Exile051112/franka-pico4-pi0-c1-red-yellow-real

Franka Pico4 pi0 LoRA

This repository contains the final PEFT LoRA adapter from the Franka/Pico4 LeRobot experiment for c1_red_yellow_real. It was trained with PyTorch LeRobot from lerobot/pi0_base for 10,000 steps.

Contents

• —adapter_model.safetensors: trained LoRA weights (about 5.6 MB)
• —adapter_config.json: PEFT configuration (r=16, lora_alpha=16)
• —config.json, train_config.json: LeRobot pi0 policy and training metadata
• —policy_preprocessor*.json and policy_postprocessor*.json: image/state/action preprocessing
• —tokenizer/: tokenizer used during training

Loading

The adapter is not a standalone full model. Download the gated base model lerobot/pi0_base with the required Hugging Face access, then load this repository as a PEFT adapter in the LeRobot pi0 policy. The adapter configuration already points to lerobot/pi0_base.

The policy expects the Franka/Pico4 contract recorded in config.json: two RGB camera streams (observation.images.top and observation.images.wrist), a 17-dimensional state, and a 10-dimensional TCP action. Validate camera names, state/action ordering, calibration, and safety limits before robot deployment.

Training

• —Dataset condition: c1_red_yellow_real
• —Precision: bfloat16 with gradient checkpointing
• —Batch size: 4
• —Steps: 10,000
• —Checkpoint: final 010000
• —Backend: PyTorch LeRobot