I am a Ph.D. student in Computer Science at the University of Maryland, College Park, working with Prof. Pratap Tokekar and Prof. Ming Lin. Previously, I earned my bachelor’s degree from Shanghai Jiao Tong University. I have also interned at Apple, Tencent AI Lab and Tencent Robotics X.
I work on Multimodal Learning, Reinforcement Learning, and Imitation Learning etc. My work spans across Multimodal Large Language Models (MLLMs) Reasoning, Robotics, and Autonomous Driving. I aim to build autonomous AI agents.
VOGUE, a method that shifts exploration to the visual input space by quantifying policy sensitivity to visual perturbations, enhances reasoning in multimodal large language models.
Vision-SR1 uses RL to enhance reasoning in vision-language models by decomposing the process into visual perception and language reasoning stages, improving accuracy and reducing hallucinations.
A RL framework that enhances LLMs reasoning capabilities by enabling parallel thinking through a progressive curriculum, leading to significant performance improvements on math benchmarks.
CDE enhances Reinforcement Learning with Verifiable Rewards (RLVR) by using intrinsic curiosity signals from the actor and critic to improve exploration and reduce premature convergence in LLMs.
A broadly applicable framework that dynamically modulates guidance signals based on associated uncertainty, providing a simple yet effective solution for incorporating uncertainty-aware guidance across diverse machine learning systems.
A multi-modal multi-agent framework that enables agents to collaborate and share multimodal data during training while allowing inference with reduced modalities during testing, which is especially beneficial for deployment in resource-constrained environments.
A multi-modal collaborative decision-making approach for connected autonomy.
A thorough iteration complexity analysis for the risk-sensitive policy gradient method, focusing on the REINFORCE algorithm and employing the exponential utility function.
A multi-dimensional Representation Learning approach that integrates visual, physical, temporal, and geometric representations to enhance the robustness and generalizability of Imitation Learning for food acquisition.
Long-horizon Visual Action-based (LAVA) food acquisition of liquid, semisolid, and deformable foods.
A data-driven technique for estimating the uncertainty distribution and bound for the KL divergence for distributionally robust optimal control (DROC).
An adaptive visual imitation learning approach that exhibits adaptability and robustness across different bowl configurations and diverse food types for robotic scooping.