Abstract:
Coordinating a fleet of autonomous robots to allocate tasks and plan paths intelligently remains one of the hardest open problems in multi-agent AI. Existing approaches either handle task allocation or motion planning in isolation and rule-based systems break down the moment the environment changes faster than the rules anticipate. AQUA Guardian tackles this by combining both layers into a single end-to-end framework. Task allocation is handled by the HybridRL agent, which integrates MAPPO, QMIX, MADDPG and DQN. Each algorithm contributes to the fnal dispatching decision and a performance-weighted selection mechanism aggregates their outputs based on cumulative task outcomes. Robots are aware of task dependencies, so no robot is sent to a job before its prerequisites are complete. Path planning operates through the HybridDiffusion model, built from TDM, DiffusionPolicy, Diffuser and SDE running in parallel. Each planner generates candidate trajectories that are scored and fused into a single smooth, collision-free path. The framework is simulated in Gazebo Harmonic and ROS 2 Jazzy running on Ubuntu 24.04, and evaluated across three scenarios of increasing complexity. These span warehouse logistics, search-and-rescue and openwater maritime patrol. Task completion rate and time to fnish capture operational performance while path quality and energy use reflect trajectory efciency. The system is built to support both ofine and online training.
