MarineFormer: A Transformer-based Navigation Policy Model for Collision Avoidance in Marine Environment

In this work, we investigate the problem of Unmanned Surface Vehicle (USV) navigation in a dense marine environment with a high-intensity current flow. The complexities arising from static and dynamic obstacles and the disturbance forces caused by current flow render existing navigation protocols inadequate for ensuring safety and avoiding collisions at sea. To learn a safe and efficient robot policy, we propose a novel methodology that leverages attention mechanisms to capture heterogeneous interactions of the agents with the static and moving obstacles and the flow disturbances from the environment in space and time. In particular, we refine a temporal function with MarineFormer, a Transformer navigation policy for spatially variable Marine environment, trained end-to-end with reinforcement learning (RL). MarineFormer uses foundational spatio-temporal graph attention with transformer architecture to process spatial attention and temporal sequences in an environment that simulates a 2D turbulent marine condition. We propose architectural modifications that improve the stability and learning speed of the recurrent models. The flow velocity estimation, which can be derived from flow simulations or sensors, is incorporated into a model-free RL framework to prevent the robot from entering into high-intensity current flow regions including intense vortices, while potentially leveraging the flow to assist in transportation. The investigated 2D marine environment encompasses flow singularities, including vortices, sinks, and sources, representing fundamental planar flow patterns associated with flood or maritime thunderstorms. Our proposed method is trained with a new reward model to deal with static and dynamic obstacles and disturbances from the current flow.