Contrastive Learning in NMPC Behavior Cloning for Generalized Trajectory Tracking of Marine Vehicles

초록

Recent studies have extensively explored representing control policies through neural networks, leveraging their universal function approximation capabilities. Imitation learning enables the replication of an expert's proficiency, significantly reducing the resources required for expert deployment. In this context, our research aims to utilize neural networks for imitating a nonlinear model predictive control (NMPC) policy, thereby reducing the computational demands for optimization. In trajectory tracking problems, while ranges of motion states, such as operating velocities can be easily determined, defining the data distribution for conceivable trajectories remains challenging. To enhance the efficiency of imitation learning and to learn a general trajectory tracking policy, we propose a method for learning a representation of trajectories using contrastive learning. By employing features learned from this method, we can identify the representative trajectories and imitate the NMPC policy with a reduced amount of demonstration data. Furthermore, the trained policy, represented as a neural network, can significantly reduce the computation time for control, benefiting small marine vehicles.

키워드