mlsanityData to 14 Oct 2026

Language-Structured Relational Q-Learning for Threat-Aware Control in Safety-Critical Driving

arXiv 2608.11498 ↗First posted 11 Aug 2026cs.CV cs.ET cs.LGcited by 0

On the map: Deep reinforcement learning methods → RL for autonomous driving and traffic

Abstract (arXiv)

Natural-language-based scenario generation offers an intuitive means of describing rare and complex driving interactions, yet it is still uncertain whether training with language-structured data leads to truly adaptive control policies. We propose Language-Structured Relational Q-Learning, instantiated through an Ego-Centric Relational Q-Network (ERQ-Net), which jointly learns inter-vehicle relevance and action values from dynamic traffic graphs. Language descriptions define surrounding-vehicle behaviours during training, while prompts and semantic actor roles are hidden from the policy. ERQ-Net must therefore infer threat relevance solely from observable kinematics and interactions. Across 2,500 safety-critical scenarios, language-structured training improves test success from 49-52% to 55-58% and increases adversary-focused attention from 1.2x to 2.1x, demonstrating emergent threat awareness. However, this representational gain does not consistently translate into adaptive control: trained policies perform similarly to the best constant action, while a portfolio of simple policies solves 76% of scenarios. We formalise this discrepancy as a recognition-control gap and show that reward reweighting and margin shaping do not eliminate the resulting policy collapse. Evaluations of realism, criticality, semantic accuracy, and transfer of state-interface representations to CARLA further highlight both the strengths and the constraints of language-structured relational policy learning in safety-critical driving scenarios.

Key facts. No citing papers in the map yet (data through 14 Oct 2026). Placed in RL for autonomous driving and traffic, one of 15 directions in Deep reinforcement learning methods. Builds on 17 of its 23 references that are in the map. First posted on arXiv 11 Aug 2026.

Builds on

17 of 23 references.

References · 17

  1. 2017Graph Attention Networkscited by 8,923
  2. 2017CARLA: An Open Urban Driving Simulatorcited by 2,449
  3. 2020VectorNet: Encoding HD Maps and Agent Dynamics from Vectorized Representationcited by 530
  4. 2021TrafficSim: Learning to Simulate Realistic Multi-Agent Behaviorscited by 193
  5. 2021Generating Useful Accident-Prone Driving Scenarios via a Learned Traffic Priorcited by 139
  6. 2021AdvSim: Generating Safety-Critical Scenarios for Self-Driving Vehiclescited by 137
  7. 2023Language-Guided Traffic Simulation via Scene-Level Diffusioncited by 111
  8. 2024ChatScene: Knowledge-Enabled Safety-Critical Scenario Generation for Autonomous Vehiclescited by 93
  9. 2023Language Conditioned Traffic Generationcited by 87
  10. 2018Scenic: A Language for Scenario Specification and Scene Generationcited by 83
  11. 2020Learning to Collide: An Adaptive Safety-Critical Scenarios Generating Methodcited by 75
  12. 2022SafeBench: A Benchmarking Platform for Safety Evaluation of Autonomous Vehiclescited by 66
5 more
  1. 2023CAT: Closed-loop Adversarial Training for Safe End-to-End Drivingcited by 54
  2. 2024CtRL-Sim: Reactive and Controllable Driving Agents with Offline Reinforcement Learningcited by 27
  3. 2024Enhancing Autonomous Vehicle Training with Language Model Integration and Critical Scenario Generationcited by 19
  4. 2025Text2Scenario: Text-Driven Scenario Generation for Autonomous Driving Testcited by 18
  5. 2024GOOSE: Goal-Conditioned Reinforcement Learning for Safety-Critical Scenario Generationcited by 10

Built on by

Not cited yet.

Similar, not cited

Close in wording, no citation either way

  1. 2025LA-RL: Language Action-guided Reinforcement Learning with Safety Guarantees for Autonomous Highway Driving
  2. 2021Towards Learning Generalizable Driving Policies from Restricted Latent Representations
  3. 2025Learning from Risk: LLM-Guided Generation of Safety-Critical Scenarios with Prior Knowledge
  4. 2025DriveMind: A Dual Visual Language Model-based Reinforcement Learning Framework for Autonomous Driving
  5. 2025SafeAuto: Knowledge-Enhanced Safe Autonomous Driving with Multimodal Foundation Models
  6. 2025VLM-SAFE: Vision-Language Model-Guided Safety-Aware Reinforcement Learning with World Models for Autonomous Driving
  7. 2022How to Learn from Risk: Explicit Risk-Utility Reinforcement Learning for Efficient and Safe Driving Strategies
  8. 2026Threat-guided Policy-aware Scene Perturbation for Safe Autonomous Driving with Online Reinforcement Learning