Taming an autonomous surface vehicle for path following and collision avoidance using deep reinforcement learning

Model-Reference Reinforcement Learning for Collision-Free Tracking Control of Autonomous Surface Vehicles

“…The optimization-based algorithms can directly find a control law with collision avoidance by optimizing a certain objective function, e.g. model predictive control (MPC) [9] and reinforcement learning (RL) [18]. They potentially have a better performance than the motion planning approaches in dynamic environments.…”

“…They will, therefore, experience dramatic degradation in performances for uncertain ASVs. In comparison to MPC, RL can learn an intelligent collision avoidance law from data samples [18], [19], which can significantly reduce the dependence on modeling efforts and thus make RL very suitable for uncertain ASVs.…”

“…In comparison to existing model-based methods, RL is capable of learning a complex tracking control law with collision avoidance from data samples using much less model information [19], [33]- [35]. It is, therefore, more promising in controlling systems subject to massive uncertainties and disturbances as ASVs [25], [36] and meanwhile achieving collision avoidance [18], given the sufficiency and good quality of collected data. Nevertheless, it is challenging for model-free RL to ensure closed-loop stability, though some research attempts have been made.…”

Model-Reference Reinforcement Learning Control of Autonomous Surface Vehicles with Uncertainties

“…In comparison with model-based methods, RL is capable of learning a control law from data samples using much less model information [22]. Hence, it is more promising in controlling systems subject to massive uncertainties and disturbances as ASVs [12], [14], [23], [24], given the sufficiency and good quality of collected data. Nevertheless, it is challenging for model-free RL to ensure closed-loop stability, though some research attempts have been made [25].…”