Empowering Autonomous Driving with Large Language Models: A Safety Perspective

TL;DR

This paper tackles safety challenges in autonomous driving caused by deep neural networks' opacity and poor generalization to long tail scenarios. It proposes integrating large language models as intelligent decision-makers for behavioral planning, coupled with a safety verifier shield that enables in-context safety learning. Two simulation case studies are presented: an LLM conditioned adaptive MPC for trajectory planning and an interactive LLM based planning framework with a state machine and memory. Results indicate improved safety metrics and driving performance over baselines, highlighting the potential of LLMs to enhance AV safety and generalizability in complex scenarios.

Abstract

Autonomous Driving (AD) encounters significant safety hurdles in long-tail unforeseen driving scenarios, largely stemming from the non-interpretability and poor generalization of the deep neural networks within the AD system, particularly in out-of-distribution and uncertain data. To this end, this paper explores the integration of Large Language Models (LLMs) into AD systems, leveraging their robust common-sense knowledge and reasoning abilities. The proposed methodologies employ LLMs as intelligent decision-makers in behavioral planning, augmented with a safety verifier shield for contextual safety learning, for enhancing driving performance and safety. We present two key studies in a simulated environment: an adaptive LLM-conditioned Model Predictive Control (MPC) and an LLM-enabled interactive behavior planning scheme with a state machine. Demonstrating superior performance and safety metrics compared to state-of-the-art approaches, our approach shows the promising potential for using LLMs for autonomous vehicles.

Figures (7)

• Figure 1: Overview of possible LLM integration for AV with a safety verifier as a shield. Most directly, LLM can make behavior-level decisions such as lane changing by scene understanding via text, which affects the trajectory planning with different safety constraints, as shown in our case studies. The safety verifier checks the safety of the proposed control input from the decision-making and conducts in-context learning if the action is verified to be unsafe, as shown in green arrows. The unsafe feedback can be traced back to the behavior maker, predictor, and perception module as shown. Besides, LLM can assist the perception module in understanding the scene for decision-making better. LLM can also help intention prediction by reading the recent history of the surroundings to better guess their driving habit and intentions (e.g., whether lane changing) for safer decision-making.
• Figure 2: This framework shows LLM as a behavior planner that provides safety constraints for a low-level MPC trajectory planner. The LLM driver takes high-level intention prediction, scenario description, behavior state machine, and its memory via text generated by a template and makes a behavior decision based on its understanding of the driving scene. LLM decisions will formulate safety constraints for low-level MPC-based trajectory planning. Serving as a verifier, the feasibility of the MPC problem will be sent back to LLM to (re)-evaluate its decision for in-context safety learning.
• Figure 3: In-context safety learning for LLM with the feedback from MPC for trajectory planning.
• Figure 4: The ego car is in blue and other agents are in yellow. The blue dots are the planned trajectory waypoints of the ego. The red dots are the sampled waypoints of other agents from the interval-based prediction. The grey rectangles are the recent trajectory histories of the ego and other agents. The LLM exhibits safe lane keeping, optimistic lane changing, cautious lane changing abort, and conservative failsafe in the simulations.
• Figure 5: Interactive multi-step decision-making based on the behavior state machine and intention prediction. The memory will store the previous observations of other agents and the reflection module will check hard constraints including state transition rules and minimum safety requirements.