Abstract
Autonomous mobile robotic systems are rapidly becoming part of our daily lives, and their use in transportation, delivery, surveillance, medical service, and household applications has been growing. With the increase in their popularity, assuring their and their surroundings' safety becomes more critical than ever. Unfortunately, these systems might be subject to various uncertainties such as unknown noises, external disturbances, intermittent sensing, model changes, and actuator faults when deployed in the real world. Such uncertainties bring challenges in motion planning as they can cause a system to deviate from its desired behavior and possibly lead to unsafe situations (e.g., an unmanned aerial vehicle (UAV) collision with an obstacle due to wind disturbance). In most cases, these uncertainties are unknown until the system starts its operation, making it difficult to train and develop techniques to increase resilience and safety during design time. Furthermore, autonomous mobile systems typically have limited computational resources on board, bringing an additional challenge at runtime. Thus, these uncertainties and their effects need to be considered in a computationally efficient way to improve the safety of autonomous systems.
This dissertation presents a set of motion planning, runtime monitoring, and validation techniques to provide safety for mobile autonomous systems operating under unforeseen uncertainties. The frameworks we develop predict the future states of the system under uncertainties and utilize these predictions to improve their safety by performing replanning accordingly while taking computational restrictions into consideration. The techniques presented in this dissertation are extensively validated through realistic simulations and on real state-of-the-art vehicles.
First, we introduce a reachability-based self/event-triggered scheduling framework to relax the traditional periodic sensor monitoring to reduce computational power. This framework aims to minimize the computational burden related to the unnecessary sensor monitoring operations while still guaranteeing safety leveraging reachability analysis. To further reduce the computational requirements of reachability analysis and enable its runtime use, especially for systems with complex dynamics, we introduce a novel Gaussian Process regression-based fast reachability framework to guarantee the safety of aerial vehicles subject to intermittent sensing problems. Moreover, we develop an assured safety monitoring and motion planning framework that leverages verified neural networks to bypass the reachability analysis computation altogether at runtime while still guaranteeing safety under unforeseen disturbances.
As the runtime disturbances can lie outside of the training bounds, unsafe situations can occur due to a lack of training data. To deal with such issues, we present an online learning framework to predict the runtime disturbances and their effects on the system's behavior to prevent unsafe situations. With this novel technique, the system can adapt its behavior under unforeseen disturbances, particularly unknown payload variations, to remain safe while improving its decision-making over time. To further deal with unforeseen system failures that cause performance degradation, we introduce a meta-learning-based framework that allows the system to predict the future states under faults and adapt its trajectory to improve tracking performance at runtime.
