Deep Learning-Based Path Planning Under Co-Safe Temporal Logic Specifications

This paper introduces an innovative deep learning framework for robotic path planning. This framework addresses two fundamental challenges: (1) integration of mission specifications defined through Linear Temporal Logic (LTL), and (2) enhancement of trajectory quality via cost function integration within the configuration space. The proposed method shows better efficacy compared to traditional sampling -based path planning methods in computational efficiency, due to its end -to -end neural network architecture. The framework functions in two key phases. Initially, using a Conditional Variational Autoencoder (CVAE), the proposed approach efficiently identifies and encodes optimal trajectory distributions. From these distributions, candidate control sequences are generated. Subsequently, a specialized neural network module selects and fine-tunes these sequences, ensuring compliance with the LTL specifications and achieving nearoptimal solutions. Through rigorous simulation testing, we have validated the effectiveness of our method in producing low-cost trajectories that fulfill LTL mission requirements. Comparative analysis against existing deep learning -based path planning methods reveals our framework's superior performance in both trajectory optimality and mission success rates.