Learning secure corridors for model predictive path following control of autonomous systems in cluttered environments

Safe, collision-free movement of autonomous systems such as robots, mobile platforms, or drones in cluttered environments is challenging. Often the exact positions and dimensions of other systems and objects are uncertain. However, data from successful, previous trajectories might be available. We design a learning-supported model predictive controller for autonomous systems to navigate through a "safe" path-corridor learned from prior collision-free movement trajectories using Gaussian process regression. The posterior mean and variance of the Gaussian process define a corridor that allows for safe transition through the cluttered environment. A model predictive controller is used to find the optimal path inside the learned corridor and steers the system to follow it. It guarantees satisfaction of constraints on the system as well as on the reference path which is subject to the learned corridor limitations. Simulation studies for an autonomous mobile robot that navigates through an environment with obstacles demonstrate the approach's benefits. It is shown that the controller's flexibility to move freely in the safe path corridor increases the performance when compared to using a predefined fixed path.