Research · Yonsei University
Real-time diffusion policies
A three-month research project at Yonsei University spanning dual-arm data collection, asynchronous diffusion-policy execution, and post-training from robot failures and human interventions.
What I worked on
- Built and operated a dual-arm 14-DoF YAM teleoperation setup, helped develop its control framework, and collected real-robot demonstrations and human interventions for policy post-training.
- Implemented real-time chunking for diffusion policies, using diffusion-space interpolation so the robot could execute one action chunk while generating the next.
- Enabled asynchronous robot control while maintaining policy performance, avoiding the degradation caused by naively averaging overlapping action chunks.
- Implemented and evaluated RL-Token-style offline post-training using failure and intervention data, contributing to top-eight finishes in both tracks of the RSS 2026 Robotics Foundation Models Challenge.