Collaborative Control for Autonomous Underwater Vehicles in Complex Marine Dynamic Environment

Conventional research on multi-AUV collaborative control often overlooks the nonlinearity of the solution space and the real physical constraints, causing policies that are effective in simulations to fail in real-world applications; concurrently, sparse reward schemes also lead to the slow learning speed. To address these issues, this article conducts research based on the MAPPO algorithm, aiming to solve the problem of multi-AUV collaborative control in complex marine environments. First, recognizing the discrepancies between real marine environments and simulation settings, we establish a simulation mechanism driven by the integration of special single vortex field and AUV dynamic models. Second, to tackle the challenge of balancing exploration and exploitation in multi-agent systems, we construct a multidimensional dynamic comprehensive reward model, effectively mitigating the curse of dimensionality present in traditional algorithms operating in high-dimensional state spaces, ensuring the convergence of the algorithm. The experimental results show that our method can complete collaborative control tasks and is robust.