Autonomous discovery of traffic laws with AI traffic scientists
06:00 · July 3, 2026 · arXiv cs.AI RSS
Universal traffic laws describe recurrent patterns in congestion, mobility and driving behavior across cities, providing a scientific basis for transportation planning, management and control. Their discovery, however, remains expert-driven, requiring candidate regularities to be identified from heterogeneous observational evidence or validated through intervention experiments. Although autonomous artificial intelligence (AI) systems have advanced scientific discovery in controlled laboratory settings, extending them to complex transportation domains remains a challenge. Here we present TrafficSci, an agentic AI system that formulates traffic-law discovery as an iterative, auditable workflow integrating evidence scoping, critic-judge hypothesis induction, and observational-interventional validation. Across four case studies spanning population, network, control and trajectory scales, TrafficSci autonomously rediscovers three established traffic laws and identifies an unreported intrinsic temporal memory scale in urban driving behavior, statistically consistent across eight cities and two trajectory datasets. TrafficSci provides a route for extending AI-driven scientific discovery from controlled domains to complex urban systems.
Summary
TrafficSci is an agentic AI system that automates the discovery of universal traffic laws—recurrent, testable regularities in congestion, mobility patterns and driving behavior that hold across cities. Traditional discovery has relied on expert-driven literature synthesis, manual variable design and repeated cycles of data cleaning and analysis, a process made harder by the open, non-stationary nature of urban traffic that precludes controlled physical experiments. TrafficSci addresses this by organizing the workflow into three interacting modules: a literature-based agent tree search that assembles an evidence corpus, a hypothesis-induction stage that formulates falsifiable candidate laws anchored to retrieved sources, and an observational–interventional validation stage that tests each candidate against real-world trajectory or network data and, where appropriate, simulation-based interventions.
In four case studies spanning population-level mobility scaling, network congestion dynamics, control interventions and microscopic driving trajectories, the system recovered three previously published traffic laws without any hand-specified hypotheses or analysis templates. Beyond rediscovery, it identified a previously unreported intrinsic temporal memory scale in urban driving behavior. The regularity proved statistically consistent across eight cities and two independent trajectory datasets, suggesting a stable characteristic time window over which drivers integrate preceding traffic conditions.
By coupling hypothesis generation directly to dual-mode validation inside a single iterative loop, TrafficSci demonstrates that LLM-based agent workflows can be extended from closed laboratory domains to complex, real-world urban systems where both observational corroboration and interventional testing are required. The resulting process is auditable and repeatable, offering a practical route to enlarge the set of interpretable scientific priors available for transportation planning and control.
Why it matters
This research is highly relevant for Dutch AI researchers and urban planners, given the Netherlands' strong focus on smart city infrastructure and advanced traffic management. The introduction of an agentic AI for autonomous scientific discovery offers actionable methodologies for institutions like TU Delft or Rijkswaterstaat to optimize urban mobility.
