Risk-Aware Decision-Making and Motion Planning for Robots in Shared Dynamic Environments

Robots operating in dynamic environments must continuously make decisions throughout task execution, across a hierarchy that ranges from high-level goal selection and route planning to low-level control actions. Real-world environments are uncertain: the time required to reach a location may vary, the position of objects may be unknown, and sensor measurements may be noisy or incomplete. A robot’s knowledge is never perfectly accurate. According to risk management, a risk can be defined as an uncertain event that, if it occurs, may compromise the task success. In robotic navigation, risk therefore extends beyond the possibility of collisions and also includes factors such as delays, inefficient trajectories, or failures to complete the task within required time limits. Identifying and explicitly accounting for these risk factors at every decision-making stage is essential to improve safe and efficient robot deployment in real-world environments. This thesis explores various techniques for managing risk across the task execution pipeline, classifying them based on their expressivity in representing risk factors and their level of mathematical formality. Rather than proposing a single unified methodology, I provide tailored strategies for addressing the risk factors that are most critical at each step of the task execution. For robotic manipulators, the research focuses on risk minimization during motion planning by leveraging the expressive power of fuzzy logic and genetic algorithms. For mobile robots, the thesis addresses the full decision-making stack, from high-level task planning to local execution. At the task level, a framework based on co-safe Linear Temporal Logic provides probabilistic timing guarantees under object location uncertainty. At the routing level, a Markov Decision Process formulation enables a tractable version of a stochastic shortest path with recourse that minimizes the risk factor of encountering humans during motion execution in warehouse environments. Finally, at the control level, a local planning method based on heterogeneous probabilistic risk maps enables safe navigation in dynamic settings.