Research map / Large language models / Retrieval and knowledge
Hybrid-Driving: An Autonomous Driving Decision Framework Integrating Large Language Models, Knowledge Graphs and Driving Rules
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Hybrid-Driving combines ChatGPT-3.5 with a TTC-based Scenario Evolution Knowledge Graph (SEKG) for action risk prediction and hand-written driving rules that filter actions, improving success rates in Highway-Env over LLM-only, DiLu and GRAD.
Problem
LLMs used as standalone autonomous driving decision-makers hallucinate, creating safety risks; existing LLM driving methods lack dynamic scenario prediction, risk assessment, and driving rules.
Approach
Discretize the observation area into 4 subareas with TTC levels to form compact scenario nodes, build a KG (SEKG) of scenario transitions with action risks, and feed rule-filtered actions plus SEKG risk values into an LLM prompt for the final decision.
Key results
- Hybrid-Driving raises success rate over LLM alone by 72.5%, 60%, and 37.5% across the three highway settings.
- Hybrid-Driving reaches SR of 80%, 72.5%, and 50% versus DiLu's 70%, 65%, and 40% (about 7.5-10 points higher) and GRAD's lower scores.
- An SEKG built only from the lane-4 setting loses just 2.5% and 5% SR when tested on lane-5 settings.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| integrate SEKG and driving rules with the LLM (ablation) | LLM autonomous-driving decisions · lane-4-density-2 | success rate (SR): 7.5% -> 80.0% | Helped |
| integrate SEKG and driving rules with the LLM (ablation) | LLM autonomous-driving decisions · lane-5-density-2.5 | success rate (SR): 12.5% -> 72.5% | Helped |
| integrate SEKG and driving rules with the LLM (ablation) | LLM autonomous-driving decisions · lane-5-density-3 | success rate (SR): 7.5% -> 45.0% | Helped |
| add SEKG without driving rules (ablation) | LLM autonomous-driving decisions · lane-4-density-2; lane-5-density-2.5; lane-5-density-3 | success rate (SR): 7.5% -> 17.5%; 12.5% -> 22.5%; 7.5% -> 15.0% | Helped |
| add driving rules without SEKG (ablation) | LLM autonomous-driving decisions · lane-4-density-2; lane-5-density-2.5; lane-5-density-3 | success rate (SR): 7.5% -> 50.0%; 12.5% -> 30.0%; 7.5% -> 15.0% | Helped |
| use SEKG built for lane-4-density-2 instead of a scenario-specific SEKG (ablation) | Hybrid-Driving · lane-5-density-2.5 | success rate (SR): 72.5% -> 70.0% | Hurt |
| use SEKG built for lane-4-density-2 instead of a scenario-specific SEKG (ablation) | Hybrid-Driving · lane-5-density-3 | success rate (SR): 45.0% -> 40.0% | Hurt |
| integrate SEKG and driving rules with the LLM | LLM autonomous-driving decisions · lane-4-density-2; lane-5-density-2.5; lane-5-density-3 | success rate (SR): 11%; 25.5%; 35% over GRAD | Helped |
| integrate SEKG and driving rules with the LLM | LLM autonomous-driving decisions · lane-4-density-2; lane-5-density-2.5; lane-5-density-3 | success rate (SR): 10%; 7.5%; 10% over DiLu | Helped |
Surprising findings
- LLM+Driving-rules outperformed LLM+SEKG, which the authors attribute to limited SEKG data.
Limitations
- SEKG suffers from inadequate data, limiting its risk prediction to basic values.
- Evaluation limited to simulated Highway-Env scenarios with ChatGPT-3.5 only.
- Driving rules are manually formulated from human expertise and not automatically learned.
Experimental setup
Datasets: Highway-Env simulation data (about 100,000 frames per scenario setting)
Benchmarks: Highway-Env lane-4-density-2 Highway-Env lane-5-density-2.5 Highway-Env lane-5-density-3
Metrics: Success Rate (SR) Success Steps (SS) mean scenario risk
Baselines: LLM alone (ChatGPT-3.5) LLM+SEKG LLM+Driving-rules DiLu GRAD
Builds on: DiLu GRAD
Closest papers by meaning
- SafeAuto: Knowledge-Enhanced Safe Autonomous Driving with Multimodal Foundation Models (ICML 2025 · similarity 0.90)
- DiLu: A Knowledge-Driven Approach to Autonomous Driving with Large Language Models (ICLR 2024 · similarity 0.89)
- ChatScene: Knowledge-Enabled Safety-Critical Scenario Generation for Autonomous Vehicles (CVPR 2024 · similarity 0.89)
- AutoVLA: A Vision-Language-Action Model for End-to-End Autonomous Driving with Adaptive Reasoning and Reinforcement Fine-Tuning (NeurIPS 2025 · similarity 0.89)
- Drive-R1: Bridging Reasoning and Planning in VLMs for Autonomous Driving with Reinforcement Learning (AAAI 2026 · similarity 0.89)
- HCRMP: An LLM-Hinted Contextual Reinforcement Learning Framework for Autonomous Driving (NeurIPS 2025 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.