On a Formal Model of Safe and Scalable Self-driving Cars

Shalev-Shwartz, Shai; Shammah, Shaked; Shashua, Amnon · 2017 · OpenAlex-citations

DOI: 10.48550/arxiv.1708.06374

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 critical challenges of safety assurance and scalability in autonomous vehicle (AV) development, arguing that current data-driven approaches are insufficient for widespread adoption. The authors contend that relying on statistical validation via miles driven is infeasible, as achieving a fatality rate three orders of magnitude lower than human drivers ($10^{-9}$ per hour) would require approximately thirty billion miles of testing. Furthermore, such statistical methods lack transparency and cannot guarantee safety in multi-agent environments where accidents may involve multiple parties. To prevent a "winter of autonomous driving" caused by unsustainable costs and lack of trust, the paper proposes a formal, white-box mathematical model for safety and a scalable system design. The primary contribution is the Responsibility-Sensitive Safety (RSS) model, which formalizes the legal concept of "Duty of Care" into rigorous mathematical rules. RSS aims to ensure that an AV never causes an accident and can compensate for reasonable mistakes by other agents, thereby achieving "AI-Utopia" where zero accidents occur if all agents follow the model. The model is built on five common-sense rules, such as not hitting vehicles from behind and respecting right-of-way. It defines "safe longitudinal distance" and "proper response" using parameters like response time and maximum braking acceleration, allowing for efficient, inductive verification of safety without requiring exhaustive simulation. Additionally, the authors introduce a semantic language for planning and sensing, which defines actions in terms of longitudinal and lateral goals rather than geometric vectors. This semantic approach bounds the computational complexity of planning and enables a Probably Approximate Correct (PAC) model for sensing that distinguishes between safety-critical and comfort-related errors. The findings demonstrate that the RSS model provides a verifiable guarantee that the AV will not be the cause of an accident, addressing the limitations of functional safety standards which do not cover logical decision-making. The semantic framework allows for offline validation using a dataset of approximately $10^5$ hours of driving data, significantly reducing the data requirements compared to statistical approaches. By defining the Q-function over a semantic space, the number of trajectories to inspect remains bounded regardless of the planning horizon, facilitating efficient machine learning. The model also supports the construction of high-definition maps through crowd-sourced, low-bandwidth data, enhancing scalability. The significance of this work lies in providing a standardized, interpretable safety model that bridges the gap between theoretical safety guarantees and practical deployment. By offering a formal definition of safe driving that is both sound and useful, the paper provides a foundation for regulatory bodies and industry stakeholders to establish consensus on AV safety. This approach ensures that autonomous vehicles can be mass-produced and deployed globally with negligible incremental costs for new cities, thereby sustaining the economic viability and societal acceptance of autonomous driving technology.

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.

Summary generated by qwen3.6-27b-prismaquant on 2026-06-26; verification: verified.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

• driverless ads