Optimizing Interaction Through Environment Design: Reward Alignment and Intent Prediction in Human-Robot Collaboration

Abstract: Effective human–robot collaboration requires accurate inference of human intent at both the task and motion levels. This thesis treats the environment as a decision variable, introducing environment design as a mechanism to improve reward alignment and intent prediction. It develops three approaches: a bilevel optimization framework for just-in-time robotic kitting, an active preference learning method that jointly optimizes environment parameters and query selection, and a quality-diversity approach for generating legible workspaces. Across user studies and simulations, these methods improve task efficiency, increase query informativeness, and enhance goal inference accuracy. Overall, the results establish environment design as a complementary paradigm to learning and inference for scalable, robust human–robot interaction.

Bio: Yi-Shiuan is a Ph.D. Candidate in Computer Science at the University of Colorado Boulder, advised by Prof. Alessandro Roncone and Prof. Bradley Hayes. His research focuses on environment design for human-robot interaction, with an emphasis on reward alignment and human motion prediction. Outside of work, he likes to run, snowboard, and play tennis.