Sophie (Ya-Chuan) Hsu

I'm a 5th-year PhD candidate at USC, advised by Professor Stefanos Nikolaidis. My research lies at the intersection of reinforcement learning (RL) and human-robot interaction (HRI).

I am interested in enabling robots to effectively collaborate with humans by inferring human's intentions and reactions in uncertain and dynamic environments. At ICAROS lab, I study diverse collaboration behaviors, develop hierarchical POMDP frameworks for human-aware robotic planning under uncertainty, and infer human observation functions to improve robot decision-making in tasks. I have interned at Toyota Research Institute working with Guy Rosman, Jonathan DeCastro and Andrew Silva on designing an assistive driving system.

Prior to joining USC, I completed my M.S. in Computer Science and Engineering at Texas A&M University, advised by Professor Dylan A. Shell. where I worked on pedestrian-aware autonomous vehicle planning. I also work with Professor Swaminathan Gopalswam and Professor Srikanth Saripall at the Texas A&M Transportation Institute, formalizing human-machine communication protocols and deploying a behavior planner using ROS on a Ford Lincoln MKZ platform.

Feel free to say hi: yachuanh at usc dot edu

CV  /  Scholar  /  Twitter

Research

This paper integrates human field-of-view (FOV) limitations into robot planning by adapting to the evolving subtask intent of humans based on their limited perception. To manage the resulting computational complexity, we propose a hierarchical online planner. In a steakhouse domain study, our FOV-aware planner reduced human interruptions and redundant actions. We further demonstrate our planner in a virtual reality kitchen environment.

We propose using surrogate models to efficiently generate diverse and reproducible failure scenarios in human-robot interaction tasks, reducing the computational cost of traditional simulation-based methods.

We propose a method using GANs and a quality diversity algorithm to generate realistic and diverse indoor furniture arrangements, varying in attributes like price and number of pieces.

Research on human-robot collaboration often focuses on robot policies for fluent teamwork, overlooking the impact of the environment on coordination. We propose a framework for procedurally generating environments that are stylistically human-like, solvable by human-robot teams, and diverse in coordination behaviors.

This paper tackles the challenge of resolving ambiguous traffic situations where autonomous vehicles cannot directly communicate intent. It proposes a model using a partially observable Markov decision process (POMDP) to produce changes in speed to express intent. The approach is validated in a simulated vehicle-pedestrian crossing and tested in real-world trials with a self-driving car, demonstrating safe and efficient navigation.

This paper is a preliminary study on communication between pedestrians and autonomous vehicles at unsignalized intersections. We propose a decision-theoretic model, using an MDP framework inspired by psychological studies, to represent pedestrian-vehicle interactions.

Service

Teaching

Inspired by this and this.