Exile051112/franka-pico4-pi0

Franka Pico4 Pi0 Models

This repository contains independent Pi0 PEFT LoRA checkpoints trained with the official PyTorch LeRobot workflow. Each condition is stored in its own folder.

Folder	Training data	Status
c1_red_yellow_real	5 red real + 5 yellow real	10k steps
c2_red_yellow_blue_real	5 red real + 5 yellow real + 5 blue real	10k steps
c3_red_yellow_blue_edited	5 red real + 5 yellow real + 5 blue edited	10k steps
c4_red_yellow_real_blue_sim	5 red real + 5 yellow real + 5 blue simulated	10k steps
c5_red_yellow_blue_sim	5 red simulated + 5 yellow simulated + 5 blue simulated	10k steps

All experiment folders contain checkpoints at 2k, 4k, 6k, 8k, and 10k steps. The adapters use PEFT LoRA (r=16, alpha=16), batch size 4, bfloat16, and gradient checkpointing. Loading requires the matching Pi0 base, tokenizer, and the selected condition folder.

The evaluation target is the blue Franka/Pico4 setup. The datasets use the robot-specific 17-D state, 10-D TCP action, top/wrist RGB streams, and 30 FPS.