Control Inference Neural Network for Motion Planning With Dynamical Systems

Abstract

This letter proposes an approach that combines the use of sampling-based motion planning with machine learning algorithms. In particular, the asymptotically optimal planner <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">SST*</i> is used to generate asymptotically minimum time trajectories, considering systems with motion constraints (nonholonomic systems) and different orders in the dynamics of the models. The intermediate states that form the trajectories obtained by the <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">SST*</i> planner, together with the controls that generated them, constitute the inputs and outputs, respectively, of the training set for a new variant of the <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">MPNet</i> network. This contrasts with other works in which neural networks are proposed that infer states that serve as a guide for another complementary method to generate controls that respect the dynamics of the system. The proposed method uses a neural network to <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">directly infer</i> the controls of the system (plus an application time of those controls), accelerating the generation time of trajectories up to three orders of magnitude in comparison to other methods based on learning. The proposed approach is evaluated with different dynamical systems and environments, including maps with concave obstacles and narrow passages.