Abstract:
Autonomous systems such as robotic systems are envisioned to co-exist with humans in our daily lives, from household service to large-scale warehouse logistics, agricultural monitoring, and smart city. Reliable interactions among robots and humans require provably correct guarantees about safety and performance when designing robot behaviors. While traditional approaches for safety and performance analysis are often used under the assumptions of perfect information about the robot models and well-defined environments, the precomputed guarantees could be easily violated when deploying robots in the real world that is uncertain, rapidly changing, and inherently stochastic. The goal of my thesis is to develop mathematically grounded algorithms that enable robots to safely and effectively interact with each other by adapting to uncertain and possibly hostile dynamic environments.

In this talk, I will first present an explicit behavior design for computationally efficient safety assurance under uncertainty on large-scale autonomous systems, such as a team of drones. In the presence of unknown robot models and uncertain environments, I will show a sample efficient safe reinforcement learning framework that integrates control-theoretic safe design into a learning-based approach for a robot to learn to optimally perform a task with safety guarantee. Next, I will talk about resilient autonomy for effective multi-robot coordination, where a cohesive group of mobile robots retains desired inter-robot information exchange capability through coordinated behaviors against defective robots. Then I will discuss how these results lead to reliable multi-robot behaviors design with guaranteed performance for practical applications such as distributed data-driven environmental sampling and monitoring. Finally, I will discuss future challenges and new ideas to build long-term autonomy that is correct by design for robots to safely and reliably collaborate with humans and each other in a variety of real-world applications.

More Information

Thesis Committee Members:
Katia Sycara, Chair
Maxim Likhachev
Changliu Liu
Amanda Prorok, University of Cambridge