Subspace Structured Neural Network for Rapid Trajectory Optimization

Abstract

In this study, finite horizon constrained trajectory optimization is tackled by using Artificial Neural Network with an embedded subspace manifold. The resulting network takes advantage of the reduced dimension search space guided by a bio-inspired motion rule. The input nodes of the network are interpreted as collocation points over the time domain transcribed by a pseudospectral discretization method. The activation function for each node is the inverse of the dynamical system. The weights and biases to be optimized in the network are analogous to the parameters of the motion rule. The network is optimized during training by minimizing an augmented loss function where the constraints are considered penalties. The proposed method is simulated in a collision avoidance trajectory planning problem of a mobile robot with two driving wheels and an attitude slewing maneuver problem of an asymmetric rigid body spacecraft.