Xinhu Li

I am an intern at USC, advised by Erdem Bıyık. Prior to this experience, I was advised by Joseph J. Lim. I received my master's degree in Computer Science from USC. Before that, I completed my B.Eng. in Software Engineering at the Zhejiang University of Technology.

Research Goal: My overarching research ambition is to create robust, versatile agents capable of navigating and resolving tasks in unstructured environments. To this end, I am drawn to several key areas: Robotics, Reinforcement Learning, Machine Learning

Email  |  CV  |  Google Scholar

Research
GRIP Teaser When a Robot is More Capable than a Human: Learning from Constrained Demonstrators
Xinhu Li, Ayush Jain, Zhaojing Yang, Yigit Korkmaz, Erdem Bıyık
In Submission

We address learning from constrained demonstrators by propagating goal proximity rewards to out-of-distribution states via confidence-based filtering and interpolation.

arXiv

SAVO Teaser Mitigating Suboptimality of Deterministic Policy Gradients in Complex Q-functions
Ayush Jain, Norio Kosaka, Xinhu Li, Kyung-Min Kim, Erdem Bıyık, Joseph J. Lim
RLC 2025, Reinforcement Learning Conference
Outstanding Paper Award on Empirical Reinforcement Learning Research

We identify that TD3 gets stuck in local optima in tasks with complex Q-functions and propose a new actor architecture to find better optima.

OpenReview | arXiv

SSL Teaser Beyond Policy Transfer: Self-Supervised Reward Adaptation
Xinhu Li*, Ayush Jain*, Joseph J. Lim, Erdem Bıyık
In Progress

we introduce a self-supervised reward adaptation method for adapting policies without human assistance. It Exceeds all other adaptation methods in manipulation and locomotion environment adaptation.

Credits to the Coolest template!