Bidirectional Planning for Autonomous Driving Framework with Large Language Model.

Area of Science:

• Autonomous Systems
• Artificial Intelligence
• Robotics

Background:

• Autonomous navigation faces challenges in dynamic environments, including safety, intent prediction, and strategic planning.
• Traditional methods exhibit limitations in adaptability and safety mechanisms for unpredictable scenarios.

Purpose of the Study:

• To introduce SafeMod, a novel framework designed to enhance safety in autonomous driving.
• To improve decision-making and scenario management capabilities within autonomous navigation systems.

Main Methods:

• SafeMod employs a bidirectional planning structure: forward planning using large language models for behavior prediction and transformer-based trajectory generation, and backward planning with Actor-Critic reinforcement learning for trajectory refinement.
• Integration of text-based environment descriptions and image data for comprehensive scene understanding.
• Utilizes policy and value functions for optimal action selection based on probability distributions.

Main Results:

• SafeMod demonstrated superior performance compared to recent planning systems on CARLA and nuScenes benchmarks.
• Significant improvements in safety and decision-making were observed in both real-world and simulation testing.
• The framework effectively integrates safety considerations and decision-making processes.

Conclusions:

• SafeMod offers a promising approach to address safety and decision-making challenges in autonomous driving.
• The bidirectional planning structure enhances adaptability and robustness in complex, dynamic environments.
• This framework has the potential to advance the reliability of autonomous navigation systems.