Neural Network Approximation Based Near-Optimal Motion Planning with Kinodynamic Constraints Using RRT

In this paper, the problem of near-optimal motion planning for vehicles with nonlinear dynamics in a clustered environment is considered. Based on rapidly exploring random trees (RRT), we propose an incremental sampling-based motion planning algorithm, i.e., near-optimal RRT (NoD-RRT). This algorithm aims to solve motion planning problems with nonlinear kinodynamic constraints. To achieve the cost/metric between two given states considering the nonlinear constraints, a neural network is utilized to predict the cost function. On this basis, a new reconstruction method for the random search tree is designed to achieve a near-optimal solution in the configuration space. Rigorous proofs are presented to show the asymptotical near-optimality of NoD-RRT. Simulations are conducted to validate the effectiveness of NoD-RRT through comparisons with typical RRT and kinodynamic RRT. In addition, NoD-RRT is demonstrated in an experiment using a Pioneer3-DX robot.