1. Measurement: Understanding mobility patterns
Accurately evaluating the current state of the transportation network and mobility patterns is the first step to improving mobility. This involves gathering and analyzing real-time and historical data from various sources to understand both current and historical conditions and trends. We need to track the effects of changes as we implement them in the network. ML powers estimations and metric computations, while statistical approaches measure impact. Key areas include:
Congestion functions
Similar to well-known fundamental diagrams of traffic flow, congestion functions mathematically describe how rising vehicle volume increases congestion and reduces travel speeds, providing crucial insights into traffic behavior. Unlike fundamental diagrams, congestion functions are built based on a portion of vehicles (e.g., floating car data) rather than all traveling vehicles. We have advanced the understanding of congestion formation and propagation using an ML approach that created city-wide models, which enable robust inference on roads with limited data and, through analytical formulation, reveal how traffic signal adjustments influence flow distribution and congestion patterns in urban areas.
Foundational geospatial understanding
We develop novel frameworks, leveraging techniques like self-supervised learning on geospatial data and movement patterns, to learn embeddings that capture both local characteristics and broader spatial relationships. These representations improve the understanding of mobility patterns and can aid downstream tasks, especially where data might be sparse or when complementing other data modalities. Collaboration with related Google Research efforts in Geospatial Reasoning using generative AI and foundation models is crucial for advancing these capabilities.
Parking insights
Understanding urban intricacies includes parking. Building on our work using ML to predict parking difficulty, Mobility AI aims to provide better insights for managing parking availability, crucial for various people, including commuters, ride-sharing drivers, commercial delivery vehicles, and the emerging needs of self-driving vehicles.
Origin–destination travel demand estimation
Origin–destination (OD) travel demand, which describes where trips — like daily commutes, goods deliveries, or shopping journeys — start and end, is fundamental to understanding and optimizing mobility. Knowing these patterns is crucial because it reveals exactly where the transportation network is stressed and where services or infrastructure improvements are most needed. We calibrate OD matrices — tables quantifying these trips between locations — to accurately replicate observed traffic patterns, providing a spatially complete understanding essential for planning and optimization of transportation networks.
Performance metrics: Safety, emissions and congestion impact
We use aggregated and anonymized Google Maps traffic trends to assess impact of transportation interventions on congestion, and we build models to assess safety and emissions impact. To build safety metrics scalably, we go beyond reactive crash data by utilizing hard braking events (HBEs). HBEs are shown to be strongly correlated with crashes and can be used for road safety services to pinpoint high-risk locations and predict future collision risks.
To measure environmental impact, we’ve developed AI models in partnership with the National Renewable Energy Laboratory (NREL) that predict vehicle energy consumption (whether gas, diesel, hybrid, or electric). This powers fuel-efficient routing in Google Maps, estimated to have helped avoid 2.9M metric tons of GHG emissions in the US alone, which is equivalent to taking ~650,000 cars off the road for a year. This capability is fundamental for monitoring climate and health impacts related to transportation choices.
Impact evaluation
Randomized trials are often infeasible for evaluating transportation policy changes. To assess the impact of a change, we need to estimate outcomes in its absence. This can be done by finding cities or regions with similar mobility patterns to serve as a “control group”. Our analysis of NYC’s congestion pricing demonstrates this method through use of sophisticated statistical techniques like synthetic controls to rigorously estimate the policy’s impact and by providing valuable insights for agencies evaluating interventions.
