Abstract
Autonomous mobile robots are becoming increasingly popular and improving the quality of our lives by revolutionizing industries such as transportation, healthcare, logistics, agriculture, emergency response, and manufacturing. This rapid advancement is driven by simulation technologies that enable comprehensive design and testing before deployment. However, the transition from simulation to reality often reveals discrepancies in modeling robots and their environments, creating challenges that can hinder smooth deployment. Additionally, even after successful deployment, robots face numerous challenges and uncertainties, including environmental changes, system aging, unexpected disturbances, and actuator faults. These challenges, which often occur at runtime without prior knowledge, can cause deviations from the systems' intended behavior, leading to unsafe conditions. Considering all these challenges, to enhance the resilience of robotic deployments and operations, it becomes critical to detect the occurrence of dynamic changes and predict the future states of the robot under such changes. Further, quickly adapting the robots’ control and planning components to align with the new dynamics is essential to recover the system and resume its intended behavior. Such adaptations not only address changes happening to the system during deployment and operation but can also be used to deal with the well-known transfer-learning problem across different robotic systems for fast and safe deployment.
This dissertation contributes to the existing state of the art in robotics operations against changes in system dynamics and failures. Through the proposed framework, future states of autonomous mobile robots are monitored to proactively update their motion plan and enhance their safety when dynamics may have changed. To this end, first, we introduce a Meta-Learning-based framework that allows the system to predict its future states and state uncertainties after dynamic changes due to unforeseen faults. This framework also monitors these predictions and the surrounding environment, allowing for real-time adjustments of the reference path to ensure safety during operations. Because this proactive path replanning focuses on safety, it may disrupt the continuation of the assigned task. To ensure that a task can resume, further adaptation of the controller and path planner is preferred for better management of the degraded system. To address this challenge, we present a conformal mapping-based transfer learning method. This approach enables the adaptation of control and path planning policies to match a reference system's behavior, bypassing the need for accurate model estimation and effectively compensating for dynamic discrepancies arising from model mismatches. These techniques are validated through extensive simulations and proof-of-concept lab experiment case studies on ground and aerial robotic systems in realistic scenarios.
