SARAD: LLM-Based Safety-Aware Hybrid Reinforcement Learning with Collision Prediction for Autonomous Driving

Ensuring both safety and efficiency in decision-making for autonomous driving systems remains a fundamental challenge. Traditional Deep Reinforcement Learning (DRL) suffers from unsafe random exploration and slow convergence, while Large Language Models (LLMs) demonstrate inherent latency in real-time inference operations. To address these limitations, this paper proposes SARAD, a novel safety-aware hybrid framework that synergizes LLMs and DRL for autonomous driving. SARAD substitutes the random exploration of DRL with Retrieval-Augmented Generation (RAG)-enhanced, LLM-guided decisions sourced from a dynamic expert knowledge repository. An attention discriminator is proposed to integrate the prior knowledge of LLMs into DRL policy optimization. A collision predictor module, fine-tuned with historical collision data, is further designed to improve vehicle safety. Extensive experiments show that SARAD achieves significant performance improvements in the Highway-Env simulator, validating the effectiveness of the proposed model in autonomous driving.

Planning Automated Driving with Accident Experience Referencing and Common-sense Inferencing

Although a typical autopilot system far surpasses humans in term of sensing accuracy, performance stability and response agility, such a system is still far behind humans in the wisdom of understanding an unfamiliar environment with creativity, adaptivity and resiliency. Current AD brains are basically expert systems featuring logical computations, which resemble the thinking flow of a left brain working at tactical level. A right brain is needed to upgrade the safety of automated driving vehicle onto next generation by making intuitive strategical judgements that can supervise the tactical action planning. In this work, we present the concept of an Automated Driving Strategical Brain (ADSB): a framework of a scene perception and scene safety evaluation system that works at a higher abstraction level, incorporating experience referencing, common-sense inferring and goal-and-value judging capabilities, to provide a contextual perspective for decision making within automated driving planning. The ADSB brain architecture is made up of the Experience Referencing Engine (ERE), the Common-sense Referencing Engine (CIE) and the Goal and Value Keeper (GVK). 1,614,748 cases from FARS/CRSS database of NHTSA in the period 1975 to 2018 are used for the training of ERE model. The kernel of CIE is a trained model, COMET-BART by ATOMIC, which can be used to provide directional advice when tactical-level environmental perception conclusions are ambiguous; it can also use future scenario models to remind tactical-level decision systems to plan ahead of a perceived hazard scene. GVK can take in any additional expert-hand-written rules that are of qualitative nature. Moreover, we believe that with good scalability, the ADSB approach provides a potential solution to the problem of long-tail corner cases encountered in the validation of a rule-based planning algorithm.