Guangyi Liu

Guangyi Liu
Postdoctoral Research Scientist
Amazon Robotics

Email: gliu[at]lehigh[dot]edu | Google Scholar | LinkedIn

Welcome

Hi, welcome to my website! I am currently a Postdoctoral Research Scientist at Amazon Robotics. I obtained my Ph.D. at the Autonomous and Intelligent Robotics (AIR) Laboratory, Department of Mechanical Engineering, Lehigh University, working under the advisory of Prof. Nader Motee. Before joining Lehigh, I obtained my B.E. degree at Beijing Institute of Technology. My research interests are risk and robustness analysis in networked control systems and perception systems. I served as a workshop organizer in ACC 2023 and organized five invited sessions at ACC 2023 - 2026.

Education

Ph.D. in Mechanical Engineering, Lehigh University, 2024
M.S. in Mechanical Engineering, Lehigh University, 2018
B.E. in Aircraft Design and Engineering, Beijing Institute of Technology, 2016

Media

“Innovative algorithms for safer robot perception” interviewed by P.C. Rossin College of Engineering and Applied Science, Lehigh University.

Selected Talks and Seminars:

(05/2025) “Distributionally Robust Multi-Agent Reinforcement Learning for Dynamic Chute Mapping” at 2nd Northeast Systems and Control Workshop (NESCW).

(09/2024) “Beyond Uncertainty: Risk-Aware Active View Acquisition for Safe Robot Navigation and 3D Scene Understanding with FisherRF” at NERC 2024.

(05/2024) “Towards Safe Learning and Perception-based Networked Control Systems: A Risk-aware Approach” at the GRASP seminar at UPenn.

(10/2023) “Distributionally Robust Mitigation of Cascading Risk in Cooperative and Competitive Vehicle Platooning” at the xLab, UPenn.

(08/2023) “Risk of Misperception in Decision-Making and Control” on behalf of my advisor Prof. Nader Motee, at the ONR Science of Autonomy Program Review.

Selected Research Areas

Distributionally Robust Multi-Agent Reinforcement Learning for Dynamic Chute Mapping (Paper: C13)
We propose a Distributionally Robust Multi-Agent Reinforcement Learning (DRMARL) framework for destination-to-chute mapping in Amazon Robotics warehouses, designed to handle uncertain and dynamic package induction rates. DRMARL integrates group distributionally robust optimization (DRO) with a contextual bandit-based predictor to improve learning efficiency and ensure robust chute mapping under varying induction conditions.

Multi-Objective Reinforcement Learning for Warehouse Optimization and Resource Allocation (Paper: P5)
We develop a Multi-Objective RL framework for large-scale tote consolidation in human-robot fulfillment centers that balances competing objectives like throughput, space utilization, and capacity constraints. Using best-response versus no-regret game dynamics, our approach learns policies that trade off multiple KPIs without manual weight tuning. We provide theoretical guarantees for extracting feasible policies and demonstrate strong performance on realistic warehouse simulations.

Beyond Uncertainty: Risk-Aware Active View Acquisition for Safe Robot Navigation and 3D Scene Understanding with FisherRF (Paper: C12, C14)
We introduce a novel approach to enhance robot navigation safety and improve 3D scene understanding by extending beyond conventional uncertainty-based methods.

Risk Analysis and Mitigation of Cascading Failures in Networked Control Systems (Papers: J1-J2 C3-C6, C8-C9)
We develop comprehensive frameworks to analyze and mitigate cascading failures across diverse networked control systems, including autonomous vehicle networks and power grids. Our work quantifies how initial failures propagate through networks under communication delays and input uncertainties, revealing fundamental trade-offs between system performance and cascading risk. We introduce data-driven distributionally robust control methods that actively mitigate cascading risks while satisfying probabilistic safety constraints, applicable to multi-agent systems and wide-area control of power networks.