Pi0.5 Post-train

Overview

This project focuses on adapting vision-language-action models to real-world robotic manipulation. I built an end-to-end VLA pipeline on a UR10e robot, covering teleoperation-based demonstration collection, multimodal data processing in LeRobot format, π0.5 post-training with LoRA, and real-time policy deployment. I further integrated RTC-based asynchronous action-chunk inference to enable continuous execution on the physical robot.

Demo

Plug RJ-45 task (original speed)

The policy was trained with Knowledge Insulation, which substantially outperformed standard LoRA fine-tuning and achieved a success rate of 16/25.

Failure case:

Pick and Place task (speed x3)

Implemented train-time RTC for the pick-and-place task.