Data-driven Learning and Control - DDLC

Abstract : This talk presents recent advances in data-driven learning and control for UAV trajectory planning in dynamic environments. We introduce three integrated approaches that leverage imitation learning to enhance real-time decision-making and robustness. Deep-PANTHER is a perception-aware planner that enables UAVs to avoid dynamic obstacles while maximizing their presence in the field of view, achieving rapid replanning through learned policies. PRIMER scales this concept to decentralized multiagent systems, where agents efficiently deconflict trajectories and handle localization uncertainties while achieving faster computation times. Finally, we discuss an approach to efficiently distill robust model predictive control policies into deep neural networks, ensuring resilience to disturbances with minimal training data. Together, these methods illustrate the power of data-driven techniques in developing scalable, efficient, and robust control systems for autonomous flight.