Obstacle Avoidance for Autonomous Sailboats via Reinforcement Learning with Coarse-to-fine Strategy

Obstacle avoidance of autonomous sailboats is a complicated task due to big inertia, highly nonlinear motion of sailboat and uncertain disturbance from wind and water current. To deal with the obstacle avoidance problem of autonomous sailboats, we promote a novel method based on reinforcement learning with coarse-to-fine strategy. In this strategy, coarse stage is used to detect the autonomous sailboat roughly in the test environment. Then fine stage is applied to localize the autonomous sailboat accurately. Hereby, the avoidance performance is improved by the transition from coarse to fine. We have verified our algorithms both in simulation and real experiments. With our method, the sailboat avoids the obstacle in higher precision than the method without the coarse-to-fine strategy. The final success rate of obstacle avoidance is near 83% and the rate of reaching goal is 70%. Both simulation results and experiments show that our method is effective.