MARL masters multi-agent rendezvous in complex fluid flows

A team led by Bocheng Li (arXiv:2606.11274) applied multi-agent reinforcement learning (MARL) to solve the rendezvous problem for swarms operating in fluid flows. In such environments, naive strategies where agents simply navigate toward each other often fail due to vortex dynamics that push agents apart or trap them in separate eddies. The MARL approach learned physics-informed coordination policies that exploit local fluid kinematics, achieving significantly higher success rates across diverse flow conditions—varying vortex intensity, scale, and number of agents. The learned policies also revealed a non-intuitive symmetry-breaking mechanism that prevents agents from becoming stuck in different vortices, allowing them to converge efficiently.

Beyond empirical performance, the authors extracted a simpler heuristic from the learned strategy that still outperformed the naive baseline. They also provided theoretical analysis showing that fluid deformation—measured via finite-time Lyapunov exponents—impedes rendezvous; regions of high deformation are where adjacent agents are most likely to be separated. This suggests that future multi-agent path planning in fluids should avoid such high-separation zones. The work highlights MARL's ability to discover swarm intelligence in complex, real-world fluid environments, with implications for underwater drone coordination, atmospheric monitoring, and biomedical micro-robotics.