[論文レビュー] On a Formal Model of Safe and Scalable Self-driving Cars

Shai Shalev‐Shwartz, Shaked Shammah+1 more|arXiv (Cornell University)|Aug 21, 2017

Reinforcement Learning in Robotics参考文献 9被引用数 420

ひとこと要約

本論文は Responsibility-Sensitive Safety (RSS) を導入し、自動運転車向けの形式かつ検証可能な安全モデルと、安全性を実用的な展開と整合させるスケーラブルなシステム設計を提案する。

ABSTRACT

In recent years, car makers and tech companies have been racing towards self driving cars. It seems that the main parameter in this race is who will have the first car on the road. The goal of this paper is to add to the equation two additional crucial parameters. The first is standardization of safety assurance --- what are the minimal requirements that every self-driving car must satisfy, and how can we verify these requirements. The second parameter is scalability --- engineering solutions that lead to unleashed costs will not scale to millions of cars, which will push interest in this field into a niche academic corner, and drive the entire field into a "winter of autonomous driving". In the first part of the paper we propose a white-box, interpretable, mathematical model for safety assurance, which we call Responsibility-Sensitive Safety (RSS). In the second part we describe a design of a system that adheres to our safety assurance requirements and is scalable to millions of cars.

研究の動機と目的

• Propose a white-box, interpretable safety model for autonomous vehicles grounded in Duty of Care (RSS).
• Define a semantic language and planning/sensing framework to enable scalable, verifiable safety.
• Show how RSS can achieve a target fatality probability of 10^-9 per hour of driving with offline data.
• Offer a system design that implements RSS in a scalable, multi-car environment.

提案手法

• Formalize RSS as a mathematical model that captures safe driving decisions in multi-agent settings.
• Introduce a semantic language for units, measurements, and action space to link planning, sensing, and actuation.
• Develop a proximally scalable planning approach using a Q-function over a semantic space with bounded trajectory considerations.
• Provide a PAC-based sensing model to handle sensing errors while adhering to RSS.
• Demonstrate star-shaped, inductive proofs for safety in longitudinal (and later lateral) scenarios with defined safe-distance and proper-response rules.

実験結果

リサーチクエスチョン

• RQ1 How can safety for self-driving cars be modeled in a way that is both sound with respect to human legal notions (Duty of Care) and verifiable?
• RQ2 Can a semantic, planning-sensing-actuation framework ensure RSS-compliant behavior at scale?
• RQ3 What parameter choices (e.g., response time, braking/acceleration limits) balance safety guarantees with traffic efficiency?
• RQ4 Is it possible to achieve the target fatality probability of 10^-9 per hour of driving using offline data and PAC-based sensing?
• RQ5 How can multi-agent safety be proven via inductive, star-shaped proofs in both longitudinal and lateral driving scenarios?

主な発見

• RSS provides a formal interpretation of Duty of Care that aims for zero accidents if all agents follow it (AI-Utopia).
• A safe longitudinal distance is derived as a function of speeds and acceleration/braking limits, enabling a provable proper response to dangerous situations.
• The model supports scalable planning with a semantic space where the number of trajectories is bounded, aiding learning and verification.
• A PAC-based sensing model is integrated to handle measurement errors while maintaining compliance with RSS.
• The approach is designed to be verifiable offline with data on the order of 10^5 driving hours, not requiring impractically large data.