Specifically, Xinshuo's research spans the tasks of object detection, multi-object tracking, re-identification, trajectory prediction, and motion planning, with an ultimate goal of building an autonomous system such as self-driving cars that can safely interact with others in multi-agent dynamic environments. Towards this goal, she develops computational models to solve each individual task using machine learning techniques such as graph neural networks, transformers, generative adversarial networks and variational auto-encoders. To make the entire robot system robust and safe, Xinshuo's research also aims to seamlessly integrate models across tasks by building differentiable pipelines, propagating uncertainties from the upstream to downstream models, and exploring the most effective structure of the differentiable pipelines.