Title:Convex Approximation of Probabilistic Reachable Sets from Small Samples Using Self-supervised Neural Networks

Abstract:Probabilistic Reachable Set (PRS) plays a crucial role in many fields of autonomous systems, yet efficiently generating PRS remains a significant challenge. This paper presents a learning approach to generating 2-dimensional PRS for states in a dynamic system. Traditional methods such as Hamilton-Jacobi reachability analysis, Monte Carlo, and Gaussian process classification face significant computational challenges or require detailed dynamics information, limiting their applicability in realistic situations. Existing data-driven methods may lack accuracy. To overcome these limitations, we propose leveraging neural networks, commonly used in imitation learning and computer vision, to imitate expert methods to generate PRS approximations. We trained the neural networks using a multi-label, self-supervised learning approach. We selected the fine-tuned convex approximation method as the expert to create expert PRS. Additionally, we continued sampling from the distribution to obtain a diverse array of sample sets. Given a small sample set, the trained neural networks can replicate the PRS approximation generated by the expert method, while the generation speed is much faster.