RI Seminar: Dr. Yulun Tian
Title: Toward Resilient and Scalable Distributed Perception: Algorithms and Systems
Dr. Yulun Tian
Massachusetts Institute of Technology (MIT)
Bio: Dr. Yulun Tian is a postdoctoral associate at the department of aeronautics and astronautics at MIT. He received the Ph.D. degree in autonomous systems from MIT in 2023 under the supervision of Prof. Jonathan How. His current research applies tools from nonlinear and distributed optimization, graph theory, differential geometry, and machine learning to develop mathematically rigorous algorithms and reliable real-world systems for multi-agent perception and navigation. His research has received several recent awards, including a 2022 Best Paper Award from IEEE Transactions on Robotics, a 2021 Honourable Mention from IEEE Transactions on Robotics, and a 2020 Honourable Mention from IEEE Robotics and Automation Letters. More about Yulun’s research can be found at tianyulun.com.
Abstract: Collaborative perception, which enables multiple agents to construct globally consistent “mental models” of the environment from local noisy observations, forms the sensory foundation of multi-agent systems in many large-scale applications. However, achieving reliable collaborative perception in the real world is hard, due to computational challenges associated with the underlying optimization problems and operational constraints imposed by practical communication networks. To address these challenges, I will present a new paradigm based on distributed Riemannian optimization that achieves resilient and scalable collaborative perception with provable performance guarantees. Motivated by applications such as large-scale collaborative simultaneous localization and mapping (CSLAM), I will discuss how the proposed framework remains resilient even under delayed communication and achieves certifiable optimality that defies the non-convexity of the original problem formulation. Building on this algorithmic framework, I will present a complete and fully distributed system for metric-semantic CSLAM. I will demonstrate the system by showing results from recent large-scale experiments where eight ground robots collaboratively map the MIT campus using only onboard sensors, computers, and communication devices. I will conclude the talk by returning to the big picture of collaborative perception and discuss opportunities for future research.
