SACA: A Scenario-Aware Collision Avoidance Framework for Autonomous Vehicles Integrating LLMs-Driven Reasoning
DOI: 10.48550/arxiv.2504.00115
archive: archived pipeline: cataloged verified
Get this paper ↗ (DOI — opens at the source; we link to it, we don't host it)
Summary
This paper addresses the challenge of reliable collision avoidance for autonomous vehicles in extreme, high-risk scenarios where traditional rule-based or reinforcement learning systems often fail. These conventional methods struggle to balance physical dynamics with legal, ethical, and emotional considerations, particularly in chaotic situations with scarce training data. The authors propose the Scenario-Aware Collision Avoidance (SACA) framework, which integrates Large Language Models (LLMs) to provide context-aware reasoning. The primary research question is how to develop a system that merges expert collision avoidance instincts with general world awareness while maintaining the low latency required for safety-critical maneuvers. The SACA framework consists of three main components: predictive scenario analysis, online LLM-driven reasoning, and scenario-preview-based deployment. The analysis module uses Hamilton–Jacobi (HJ) reachability analysis, approximated via neural networks trained on over 31 million real-world samples, to assess obstacle risk. It also employs a hybrid LSTM-CRF model to predict the motion intentions of surrounding traffic participants. These inputs form a structured prompt for the LLM. To ensure reliability, the authors fine-tuned the GPT-4o-mini model using 235,284 tokens derived from 50 expert-annotated collision cases. To address latency issues, the system uses a preview mechanism that caches strategies based on scenario similarity (above 90% similarity retrieves cached policies; below 70% triggers online reasoning) before the critical decision time. Experimental validation was conducted using a full-scale rear-wheel drive electric vehicle equipped with steer-by-wire and brake-by-wire systems. The study compared SACA against two baselines: an LLM without fine-tuning and an imitation learning approach. In a high-risk intersection scenario involving a truck running a yellow light and crossing pedestrians, the fine-tuned SACA model correctly selected a T-type drift maneuver to the right. This action redirected impact forces to the vehicle’s energy-absorbing structures, protecting the occupant cabin and avoiding the pedestrians. In contrast, baseline models either misunderstood spatial constraints or selected inappropriate maneuvers. The results demonstrated that SACA effectively reduced collision losses in extreme scenarios and lowered false triggering rates under complex conditions compared to baseline methods. The significance of this work lies in demonstrating a feasible pathway for integrating LLM-driven reasoning into safety-critical autonomous driving tasks. By combining domain-specific fine-tuning with a memory-based retrieval system, SACA overcomes the latency and robustness limitations typically associated with LLMs in real-time control. The framework successfully incorporates broader societal norms and ethical considerations into decision-making, offering a more transparent and adaptable alternative to black-box neural networks. This approach enhances the ability of autonomous vehicles to handle rare, messy situations where training data is limited, thereby improving overall road safety.
Key finding
Real-vehicle tests show that the SACA framework effectively reduces collision losses in extreme high-risk scenarios and lowers false triggering under complex conditions compared to baseline methods.
Methodology
on_road
Provenance
The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed. Discovered via author_sweep_intake on 2026-05-28.
| Stage | Outcome | Tool | Model | Prompt | Attempts | Completed |
|---|---|---|---|---|---|---|
| discover | success | author_sweep | — | — | 2 | 2026-05-28 |
| archive | success | canonical_url | — | — | 1 | 2026-06-04 |
| extract | success | cached | — | — | 3 | 2026-06-10 |
| clean | success | clean | — | — | 1 | 2026-06-04 |
| chunk | success | chunk | — | — | 1 | 2026-06-04 |
| embed | success | embed | Qwen/Qwen3-Embedding-8B | — | 1 | 2026-06-04 |
| enrich | success | — | — | — | 1 | 2026-05-28 |
| promote | success | — | — | — | 1 | 2026-06-04 |
| summarize | success | llm | qwen3.6-27b-prismaquant | summ-v5 | 2 | 2026-06-10 |
| tag | success | vector_similarity | — | — | 15 | 2026-06-11 |
| verify | success | — | — | — | 2 | 2026-06-10 |
Summary generated by qwen3.6-27b-prismaquant on 2026-06-10; verification: verified.
Topics
Ranked by relevance to this paper. Hover a topic for its definition.
Information type
What kind of knowledge this paper contributes, grouped by family — independent of topic (what it is about) and method (how it was studied).
- Theoretical Contribution: computational model