Underwater robots track targets with AI breakthrough

Underwater vehicles tracking maneuvering targets as a team is hard: acoustic communication is constrained, network topology shifts, and ocean currents add uncertainty. To tackle this, researchers propose VGG-MADiffRL, a value-gradient-guided multi-agent diffusion reinforcement learning algorithm, paired with MDCA, a diffusion-based hierarchical control architecture. MDCA is a three-tier closed-loop framework—global intelligent control, local online training, and physical action execution—that coordinates task allocation, local decisions, and execution feedback. VGG-MADiffRL uses value gradients to guide action generation during the reverse denoising process, pushing actions toward higher expected returns, while twin value networks with joint optimization and soft target updates help stabilize training. The approach models sonar detection and ocean current disturbances, framing cooperative tracking as a Markov decision process. In experiments, it delivers faster convergence, higher tracking accuracy, and smoother training dynamics in cooperative tracking scenarios, showing real promise for dynamic underwater missions.