Published in IEEE Access, 2026
We developed a human-VLM collaborative framework for preference labeling in reinforcement learning. It uses label consistency to estimate VLM reliability and Bayesian optimization to select camera viewpoints, improving annotation quality while reducing the need for human feedback.
Published in arxiv preprint, 2023
We developed a vision-and-language manipulation system that lets users assign unique names to objects before giving instructions. By remembering both the name and appearance, the system reduces ambiguity in later pick-and-place commands; in experiments, it improved task success by 31% under ambiguous instructions.
Published in 2021 IEEE International Conference on Robotics and Automation (ICRA), 2021
We developed an uncertainty-aware nonlinear model predictive controller that adjusts how strongly a companion robot trusts predicted human motion. This enables proactive following while avoiding overconfident movements that could obstruct the person or cause the robot to fall behind.
Published in Robotics and Autonomous Systems, 2020
We developed an adaptive nonlinear model predictive controller for a two-wheeled companion robot. By adjusting its cost-function weights according to the human-robot distance, the robot can follow people smoothly through situations such as U-turns without obstructing them.
