CASE STUDY · AUTONOMY RESEARCH
Safe Multi-UAV Coordination with Deep Reinforcement Learning
Doctoral research on decentralized multi-agent coordination under partial observability and explicit safety / temporal-spatial constraints.
Problem
Multi-UAV coordination becomes difficult when each vehicle has incomplete information, agents must act concurrently, and safety cannot be treated as a soft preference. Learning must produce useful cooperative behavior without ignoring the geometry and constraints of the physical system.
Constraints
My doctoral work focused on decentralized coordination with partial observability, safety requirements, and temporal-spatial constraints. The central tension was familiar to anyone combining learning and controls: how do we preserve the adaptability of reinforcement learning while making safety structure explicit?
Approach
The research explored multi-agent deep reinforcement learning methods for coordinated UAV behavior, including partial observability and safety-aware formulations. The work connected reinforcement-learning policy design with constraint handling and control-oriented reasoning.
A decentralized formulation can be viewed as each agent selecting an action from local information,
\[a_i \sim \pi_i(a_i \mid o_i),\]while the joint system must satisfy safety and coordination requirements that may depend on the relative state of multiple agents.
Results
The research produced a Ph.D. dissertation and peer-reviewed publications in AIAA, IEEE Access, and the multi-agent robotics literature. It also became a foundation for later technical work and internal reinforcement-learning education.
Engineering notes
The enduring lesson is that autonomy needs both decision intelligence and constraint intelligence. Learned policies are powerful, but engineering confidence rises when the design makes safety, observability, timing, and failure modes first-class parts of the problem statement.
See the research and publications page for citations and the dissertation link.