A Forward Propagation Motion Planning Algorithm Based on Generative Model

The random-forward-propagation-based approach can solve the kinodynamic motion planning(KMP) problem without requiring solving the boundary value problem (BVP). This technique ensures probability completeness and nearly asymptotic optimality by randomly propagating nodes and uti-lizing appropriate node selection methods. Random sampling control can help find a feasible solution, while it also generates numerous unnecessary extensions that reduce computational efficiency. In this paper, we propose a forward propagation motion planning method based on a generative model. The motion sequence is no longer entirely dependent on random generation but instead through the neural network heuristic, resulting in a faster solution of high quality. Specifically, a VAE-GAN is employed as the generative model in our approach. The shared generator in both VAE and GAN generates a set of control candidates simultaneously, from which the discriminator selects the optimal one for propagation. A large number of simulation experiments are conducted on different environments to verify the effectiveness of our algorithm.