Mission-Conditioned Path Planning with Transformer Variational Autoencoder

Abstract

This paper introduces a novel deep learning framework for robotic path planning that addresses two primary challenges: integrating mission specifications defined through Linear Temporal Logic (LTL) and enhancing trajectory quality via cost function integration within the configuration space. Our approach utilizes a Conditional Variational Autoencoder (CVAE) to efficiently encode optimal trajectory distributions, which are subsequently processed by a Transformer network. This network leverages mission-specific information from LTL formulas to generate control sequences, ensuring adherence to LTL specifications and the generation of near-optimal trajectories. Additionally, our framework incorporates an anchor control set—a curated collection of plausible control values. At each timestep, the proposed method selects and refines a control from this set, enabling precise adjustments to achieve desired outcomes. Comparative analysis and rigorous simulation testing demonstrate that our method outperforms both traditional sampling-based and other deep-learning-based path-planning techniques in terms of computational efficiency, trajectory optimality, and mission success rates.