Resilient Task Sequence Planning for Industrial Mobile Robots

This thesis addresses the development of a resilient task sequence planning framework for industrial mobile robots using reinforcement learning. The thesis will focus on how best to enhance the adaptiveness and efficiency of robots in dynamic and, subsequently, unpredictable industrial environments. In the proposed framework, Markov Decision Processes are to be used for modeling the environment in which the robot makes a decision. In contrast, several RL algorithms like Q-learning, SARSA, and Dyna-Q are proposed to optimize navigation and task execution. The key contributions of this thesis are the proposition of new models for efficient robot navigation, implementation of robust RL algorithms in Python, and a simulation framework that could validate the resilience of the robot against worst-case scenarios. It also contains a comparison between different approaches to reinforcement learning, underscoring sample-based planning methods joined with hyperparameter tuning to achieve optimal policies. Simulation results show that the proposed methods perform very well in improving the performance of the robot. Cumulative rewards indicate that successful learning and optimization of policies have occurred. The finding is that increased reliability and productivity in industrial mobile robots would enable these robots to handle real-world applications with complex tasks, recover from failures, and be based on RL for navigation and task management.