MSc thesis · UPC × ICM-CSIC 2025 — 2026
Multi-agent RL for marine hazard mapping
How can teams of autonomous surface vehicles track the boundary of a harmful algal bloom or an oil spill — with only local sensing and limited communication? This thesis introduces a flexible simulation framework with static and dynamic contamination scenarios, and evaluates TransfQMix-based decentralized policies across observation modes, communication loss, sensor noise and team-size transfer.
Findings § observed
Emergent coordination
Agents learn cooperative perimeter coverage, zig-zag traversal along the boundary, and collision-aware behavior in complex plume geometries.
Robustness limits
Policies stay robust under communication loss and sensor noise, while transfer between static and dynamic environments remains challenging — documented as a finding.
Deployment feasibility
An initial analysis of computational feasibility for edge inference shows the policies can run on constrained onboard systems.
Episode renders § in motion
Static plume scenario
Dynamic plume scenario
Explore § full record