通过神经符号安全卫士引导端到端自动驾驶

文章背景与核心概要

现代端到端自动驾驶系统在平均性能指标上表现出色，但往往难以遵守人类驾驶员能够轻松遵循的基本交通规则。这种局限性源于其依赖统计模式识别而非确定性物理定律的本质，导致决策过程不透明且难以强制执行安全约束。

为了解决这一问题，研究人员引入了一种“神经符号安全卫士”（neuro-symbolic safety guard）。这是一个轻量级的非学习型模块，被附加在预训练驾驶代理的最终输出接口上。该模块在指令到达车辆前进行实时拦截，根据显式的安全规则验证动作，并在必要时将不安全的指令替换为最近的安全替代方案。

在 Fail2Drive 和 Bench2Drive 基准测试中，基于 TransFuser v6 (TFv6) 模型的实验表明，该方法在无需重训练基础模型且不降低整体驾驶性能的前提下，成功将驾驶成功率提升了 15%，并将安全关键型碰撞事故减少了高达 53%。

执行摘要

现代端到端自动驾驶系统在平均性能指标上表现出色，但往往无法遵守基本的交通规则。这种局限性源于它们依赖统计模式识别而非确定性物理定律，使得其决策过程不透明，且难以强制执行安全保障。

Modern end-to-end autonomous driving systems excel at average performance metrics, yet frequently fail to obey fundamental traffic rules. This limitation arises from their reliance on statistical pattern recognition rather than deterministic physical laws, making their decision-making opaque and safety guarantees difficult to enforce.

为了弥补这一差距，研究人员引入了一种“神经符号安全卫士”——这是一个轻量级的、非学习型模块，附加在预训练驾驶代理的最终输出接口上。该模块在指令到达车辆前立即生效，根据显式的安全规则验证动作，并有选择地用最近的安全替代方案覆盖不安全的指令。在 Fail2Drive 和 Bench2Drive 基准测试中，使用 TransFuser v6 (TFv6) 模型进行的测试表明，该方法成功将驾驶成功率提高了 15%，并将安全关键型碰撞减少了高达 53%，且无需重训练基础模型，也不会降低整体驾驶性能。

To bridge this gap, researchers introduced a neuro-symbolic safety guard—a lightweight, non-learning module attached to the final output interface of a pre-trained driving agent. Acting immediately before commands reach the vehicle, the module verifies actions against explicit safety rules and selectively overrides unsafe commands with the nearest safe alternative. Tested on the Fail2Drive and Bench2Drive benchmarks using the TransFuser v6 (TFv6) model, this approach successfully improves the driving Success Rate by 15% and cuts safety-critical collisions by up to 53%, all without retraining the base model or degrading overall driving performance.

论文元数据

摘要

现代端到端驾驶代理可以实现较高的平均性能，但仍会违反人类驾驶员绝不会忽视的基本交通规则。其原因是结构性的：它们学习的是统计模式，而不是保证安全驾驶的物理条件，这使得它们的决策过程不透明，且安全约束无法得到强制执行。我们引入了一种神经符号安全卫士，这是一个轻量级模块，附加在已训练代理的最终指令接口上。在指令到达车辆之前，它会根据显式的安全规则检查指令，并仅在必要时将其替换为最近的安全替代方案。每次干预都是可直接执行的，并可追溯到触发它的规则，而卫士本身不需要重训练，也不增加任何学习组件。以最先进的 TransFuser v6 (TFv6) 为案例研究，在长尾基准测试 Fail2Drive 和 Bench2Drive 上进行评估，该卫士将成功率提高了 15%，并将安全关键型碰撞减少了高达 53%，同时保持了原始的驾驶评分。

Modern end-to-end driving agents can achieve high average performance yet still violate basic traffic rules that a human driver would never miss. The reason is structural: they learn statistical patterns rather than the physical conditions that guarantee safe driving, leaving their decision-making process opaque and safety constraints unenforced. We introduce a neuro-symbolic safety guard, a lightweight module that attaches to the final command interface of an already-trained agent. Immediately before a command reaches the vehicle, it checks the command against explicit safety rules and, only when necessary, replaces it with the nearest safe alternative. Each intervention is directly executable and traceable to the rule that triggered it, while the guard itself requires no retraining and adds no learned component. Evaluated on the long-tail benchmarks Fail2Drive and Bench2Drive using the state-of-the-art TransFuser v6 (TFv6) as a case study, the guard improves Success Rate by 15% and reduces safety-critical collisions by up to 53%, while preserving the original Driving Score.

关键发现与性能亮点

• 无需重训练： 严格作为现有预训练神经网络之上的后处理护栏层运行。
• 可追溯的干预： 每次安全覆盖都明确关联到一条被破坏的符号安全规则，提高了系统的透明度。
• 显著的安全收益：
  • 在严格的长尾基准测试中，整体成功率提升了 15%。
  • 安全关键型碰撞减少了高达 53%。
• 保持效用： 在不降低性能的情况下，保留了基础模型的原始驾驶评分。

• Zero Retraining Required: Operates strictly as a post-processing guardrail layer atop existing, pre-trained neural networks.
• Traceable Interventions: Every safety override is explicitly linked back to a broken symbolic safety rule, improving system transparency.
• Substantial Safety Gains:
• Boosts overall Success Rate by 15% on rigorous long-tail benchmarks.
• Reduces safety-critical collisions by up to 53%.
• Maintained Utility: Preserves the base model's original Driving Score without performance degradation.

访问与资源

• 全文论文： 查看 PDF | HTML 版本
• 源代码： TeX 源码
• 外部引用与工具：
  • Google Scholar
  • Semantic Scholar
  • NASA ADS

• Full-Text Papers: View PDF | HTML Version
• Source Code: TeX Source
• External Citations & Tools:
• Google Scholar
• Semantic Scholar
• NASA ADS

许可图标参考：