HIL-UMI Method Enhances Vision-Language Models for Robotics

A team led by Hao Dong at Peking University, in collaboration with PrimeBot, has introduced HIL-UMI, a method for collecting correction data using a handheld gripper instead of a robotic arm. This approach allows for faster and more efficient data collection, as cameras record the scene while the operator performs the task. HIL-UMI enhances the learning process by reducing the time spent on robot resets, utilizing a Universal Manipulation Interface for data gathering. The system demonstrates higher efficiency in data collection, achieving superior results compared to the traditional HG-DAgger method, reflecting a significant performance improvement.