How should autonomous vehicles drive? Policy, methodological, and social considerations for designing a driver

The paper addresses the core question of what constitutes proper driving behavior for autonomous vehicles and how such behavior should be specified, validated, and governed. AVs merge vehicle manufacturing with the decision-making traditionally performed by human drivers, making behavior design central to safety, legality, and utility. Existing regulatory frameworks harmonize vehicle systems but generally treat driving behavior as a driver-level responsibility, often set locally (e.g., state or municipal rules of the road). This creates a gap for AVs: translating locally variable, judgment-laden rules of the road into precise, machine-implementable specifications. The study explores the feasibility and implications of using rules of the road to define AV behavior, examines emerging policy efforts to formalize behavior, and investigates how formalization may reshape road user interactions and driving culture. The work aims to illuminate technical and policy pathways for comprehensive behavioral specifications that meet societal expectations for safety, lawfulness, and efficiency.

The paper surveys regulatory and standardization efforts shaping AV driving behavior: (1) Policy frameworks—In the U.S., NHTSA encourages assessment of AVs obeying traffic laws while deferring rules of the road (ROTRs) to states; some states (e.g., Nevada) require compliance. The Uniform Law Commission recommends evidence of compliance with traffic laws while recognizing flexibility. Internationally, Germany acknowledges necessary human-like judgment in exceptions; Austria requires adherence to national ROTRs; Ontario mandates compliance with its Highway Traffic Act. The UK considered a “Digital Highway Code,” with law commissions advocating a forum to adapt ROTRs for AVs. Singapore’s LTA TR 68 provides a detailed framework acknowledging ROTR translation challenges and specifying contexts where human-oriented provisions (e.g., mirrors) may not apply to AVs. UNECE Working Party 29 has issued precise behavioral specifications for certain Level 2/3 features (e.g., following distance, lane change clearance, lateral acceleration limits), highlighting the distinction between legal ROTRs and formal behavioral specifications. (2) Industry and standards—AV consortia and ISO technical reports recognize the need for machine-interpretable traffic rules and hierarchical rule structures to produce lawful, collision-free plans. IEEE P2846 defines assumptions about other road users’ behavior to support safe AV decisions. The literature also proposes safety and criticality metrics (e.g., rulebooks, KPI-based approaches, instantaneous safety metrics) and formal methods for encoding and monitoring rules, though many do not explicitly target ROTR compliance. (3) Socio-technical perspectives—Social science research indicates that law, culture, and enforcement historically privilege motorized users, with concepts like jaywalking shaping norms. Formalizing AV behavior may increase predictability but risks “hardening” behaviors, potentially altering road user hierarchies and local driving cultures. Overall, prior work underscores the need for precise, implementable specifications, governance processes, and consideration of broader social impacts.

The authors conducted an experiment to explore processes and methods for deriving formal rules from ROTRs. Two independent teams (Motional and Kontrol) formalized two Nevada statutes: (1) NRS 484B.250 (Yielding at intersections) and (2) NRS 484B.413 (Use of turn signals). The teams used a shared template to structure formal rules: (a) Rule intent and source (legal basis and safety/mobility rationale), (b) Rule scope (conditions under which the rule applies), and (c) Rule formulation (logical/mathematical statement defining satisfaction or violation, optionally with a violation metric to quantify degree of non-compliance). For yielding, the teams needed to define concepts such as right-of-way determination, conflict section (spatial area of interaction), and temporal/spatial thresholds, potentially requiring trajectory prediction. For turn signals, the statute specifies minimum signaling distances (≥100 ft in business/residential districts; ≥300 ft otherwise) prior to changing course, necessitating definitions of turn onset, direction consistency, and contextual classification (district type), as well as handling feasibility constraints for online verification with limited trajectory horizons. Each team worked independently, with Motional emphasizing the legal intent and broader safety goals, and Kontrol adhering closely to statutory text, reflecting different primary use cases (general behavior specification vs. online verification).

- Independent translations diverged despite overlaps in mathematical frameworks, revealing substantial interpretive latitude even for seemingly straightforward ROTRs. - Motional emphasized capturing both the letter and intent of the law, producing broader, more restrictive specifications that integrate ROTR compliance with other driving objectives; Kontrol adhered narrowly to statutory text, assuming other sections cover unstated aspects, yielding more specific, localized rules. - Use case and system constraints shape rule formulations: Kontrol’s online verification focus led to designs mindful of run-time performance and limited prediction horizons, influencing what could be verified in real time. - Yielding (NRS 484B.250): Compliance depends on defining right-of-way, the conflict section’s spatial extent, and temporal/spatial safety parameters; practical evaluation may require trajectory prediction and parameter choices (e.g., post-encroachment time), for which no standardized methods exist. Different parameterizations can flip compliance determinations for the same scenario. - Turn signals (NRS 484B.413): The statute’s minima (≥100 ft in business/residential, ≥300 ft otherwise) raise challenges for online verification (needing start/end of signaling and turn), context determination (district type), direction consistency, and edge cases (short road segments where prescribed signaling distances are impossible). The law does not specify maxima, and potential conflicts may arise with other statutes (e.g., restrictions on traveling long distances with a signal active). - Ambiguities and conflicts among ROTRs are common; a single, machine-precise formalization is unlikely without additional principles, violation metrics, and prioritization schemes to trade off rules when they conflict. - Findings reinforce the need for technical frameworks (formalization methods, safety concepts external to ROTRs) and governance mechanisms (forums for shared interpretation) to harmonize expectations and improve safety and efficiency.

The findings show that relying on natural-language ROTRs alone is insufficient to produce unambiguous, implementable AV behavioral specifications. The experiment demonstrates how differing interpretations, system constraints, and objectives can yield materially different formal rules, directly affecting safety, legality, and efficiency outcomes. By highlighting the need for explicit definitions (e.g., conflict zones, right-of-way), contextual information (e.g., district classification), and verification horizons, the study bridges the gap between policy expectations and engineering realities. Beyond engineering, formalizing behavior may alter interactions among road users: increased predictability can improve throughput and safety but may also harden behaviors and shift power dynamics, particularly affecting vulnerable road users. Thus, technical specification and social governance must co-evolve. Establishing shared technical and political processes—forums to align interpretations, prioritize values, and resolve conflicts—can harmonize behavioral expectations, facilitate validation and certification, and support broader societal goals for safe and efficient mobility.

This work argues for comprehensive AV behavioral specifications that translate ROTRs into precise, machine-interpretable rules while accommodating real-world constraints and societal values. Through a two-team formalization study on yielding and turn signals, the authors show that ambiguity, context dependence, and differing objectives lead to divergent specifications, underscoring the need for: (1) improved drafting of ROTRs to reduce subjectivity and clarify intent; (2) formal methods, violation metrics, and prioritization frameworks to manage conflicts; (3) integration of external safety concepts to guide parameter choices; and (4) new technical and political structures (e.g., multi-stakeholder forums) to align interpretations and values across jurisdictions and industry. Future research should develop standardized methodologies for rule translation and verification (including online verification under limited horizons), explore safety-informed parameterization (e.g., proximal safety measures), and create governance processes that democratize input, enabling consistent, location-aware behavior that benefits all road users.

The empirical study is limited in scope: two teams formalized only two Nevada ROTRs, which may not capture the full diversity of statutes, contexts, and interpretations. No quantitative performance evaluation of alternative formalizations is provided, and generalizability beyond the selected rules and jurisdictions is constrained. Practical constraints (e.g., online verification horizons, available perception/map data, trajectory prediction accuracy) influence rule feasibility and were not exhaustively addressed. The analysis acknowledges potential statutory conflicts and ambiguities but does not resolve them, emphasizing the need for broader, collaborative standardization and governance.