Real-time path planning in dynamic environments using LSTM-augmented A∗ search

Document Type

Article

Publication Date

Summer 7-21-2025

Abstract

This paper presents a novel predictive heuristic framework for simulated real-time path planning in dynamic environments, integrating Long Short-Term Memory (LSTM) neural networks, Kalman filtering, and the A∗ search algorithm. The proposed LSTM-Augmented A∗ method uses historical obstacle trajectories to predict future positions, significantly reducing the computational overhead associated with frequent re-planning and enhancing collision avoidance. To robustly handle prediction uncertainties and measurement noise, an adaptive Kalman filter is integrated alongside the LSTM predictions, forming a hybrid prediction model. The entire study—including obstacle modeling, hybrid prediction methods, path-planning algorithm implementation, and performance validation—is conducted using MATLAB® and its Deep Learning Toolbox simulations. Initially, extensive simulations in synthetic environments are used to evaluate the framework’s responsiveness to complex spatiotemporal obstacle dynamics. Subsequently, rigorous validation is performed using established benchmark simulation maps, notably Berlin_0_256.map, which represent realistic complexities and noisy conditions. Simulation results demonstrate substantial improvements in path efficiency, computational speed, prediction accuracy, path smoothness, and safety metrics. The proposed integration of LSTM predictions and Kalman filtering within the A∗ heuristic enables proactive, near-optimal path generation while preserving theoretical guarantees of admissibility and completeness. These findings underline the robustness and practical applicability of combining deep-learning predictions with classical heuristic methods in simulated dynamic path-planning scenarios.

Link to Full Text

Share

COinS