Adaptive bidirectional planning framework for enhanced safety and robust decision-making in autonomous navigation systems

Autonomous navigation systems face significant challenges in dynamic and complex environments, particularly in ensuring safety, predicting intent, and strategic planning. Traditional methods often fall short due to rigid architectures, limited safety mechanisms, and inadequate intent analysis. To address these limitations, intelligent adaptive solutions powered by large language models (LLMs) have emerged as a promising approach to meeting the multi-faceted demands of autonomous navigation. In this study, a modular and safety-oriented framework for navigation is introduced, leveraging LLMs to enhance decision-making and adaptability. The proposed framework consists of four key components: (i) an Environment Perception Module employing Bird’s Eye View (BEV)-based architectures to model the surrounding environment; (ii) a Trajectory Generation Module to produce reference trajectories and evaluate uncertainty, triggering bidirectional planners for enforcing safety constraints; (iii) a strategic integration of LLM inference to handle ambiguous situations; and (iv) seamless incorporation of long-term safety considerations into real-time operations. Unlike static, rule-based systems, the proposed framework offers a flexible and adaptive solution to complex navigation tasks, outperforming conventional approaches in safety and robustness.