AI in transportation is already running on real streets. Adaptive signal systems retime junctions from camera counts, transit apps forecast bus delays from live car traffic, fleets schedule repairs from diagnostic data, and driverless ride-hailing operates on public roads. Field measurements show real gains. They are also narrower than the headlines suggest: Pittsburgh’s Surtrac pilot cut average travel times by 25% and wait times by 40% against the signal plans it replaced, in a before-and-after drive-through study of nine intersections soon after installation, while the 34% crossing-efficiency gain reported for camera-based signal timing comes from a simulation. Data quality, real-time performance and data privacy are recognised constraints on whether these systems work.
For other areas of emerging AI, see our overview of AI in aspirational and future-tech.
AI in transportation examples, and which results come from the field
The first question to ask of any figure in this space is where it was measured. A before-and-after drive-through study on live roads, a vendor’s early numbers and a simulation run in Pygame are three different kinds of evidence.
| Application | Example | Reported result | Kind of evidence |
|---|---|---|---|
| Adaptive signal control | Surtrac, nine intersections in East Liberty, Pittsburgh, from June 2012 | 25% lower average travel time, 40% lower average wait time | Field before-and-after study |
| Adaptive signal control (all types) | FHWA Every Day Counts programme | Travel time improved by more than 10% on average, 50% or more where timing was badly outdated | Federal programme summary |
| Signal timing recommendations | Google Green Light, launched 2023, over 100 cities | Up to 30% fewer stops and 10% lower greenhouse gas emissions | Vendor’s “early numbers” |
| Camera-based signal timing | YOLOv3 vehicle counting feeding an adaptive timer | About 34% better crossing efficiency than static lights | Simulation only |
| Bus delay prediction | Google Maps live bus delays, hundreds of cities | Within 1.5 minutes of real bus motion on one Sydney trip | Single worked example in Google’s blog |
| Ride-hailing ETA | DiDi | ETA error reduced to 11% MAPE | Industrial report cited in a 2020 survey |
| Driverless ride-hailing | Waymo rider-only service, 56.7 million miles to January 2025 | Statistically significant lower crash rates than human benchmarks | Crash-rate comparison against human benchmarks |
A Wisconsin DOT research report gives a useful corrective to the headline numbers: it cites traffic flow improvements of up to 6% in some projects, with AI operations implementation costs ranging from $300,000 for research projects to $1 million for larger implementations. The same report found that respondents rated asset management and operations as the most ready domains, with digital twins and generative AI further out.
How does AI improve real-time traffic management?
AI improves real-time traffic management by replacing fixed signal plans with timings computed from what is actually arriving at the junction, and by predicting congestion early enough to act on it. The loop has three parts: detect and count vehicles, estimate traffic density from the count, then decide how long each green light stays on.
Detection
Surtrac uses cameras on signal masts to observe incoming traffic, and the authors note that video detection added no cost beyond what vehicle-actuated control already required. Inductive loops, radar, Bluetooth detectors and cameras also serve traffic monitoring. The Surtrac authors name detection as one source of error: lane changes cause miscounts, turning trucks and stopping buses block lanes, and side streets add flows nobody is measuring.
Deciding the timing
There is more than one way to do the deciding. Surtrac takes a decentralised, schedule-driven approach: each intersection computes a near-optimal schedule over a rolling horizon and passes its projected outflows to its neighbours so they can coordinate. Simpler camera systems count vehicles per approach and lengthen green time for busier directions within minimum and maximum caps. Research has also explored combining large language models with reinforcement learning for signal control policy. FHWA lists established adaptive systems such as SCOOT and SCATS among the available technologies.
What real deployments teach
Surtrac’s operators reported that limitations from uncertainty, disruptions and multi-modal demand became apparent only through continued operation, and they added strategies for queue estimation and bus dwell times afterwards. A USDOT benefit record estimates a benefit-cost ratio of almost 20:1 after five years if the system were deployed city-wide in Pittsburgh, which is a projection, not a measured outcome. Green Light takes a different route entirely: it infers existing signal parameters from Google Maps driving trends and hands recommendations to the city, which implements them on its existing equipment.
Adoption remains limited. FHWA states that adaptive signal control has been deployed on less than 1% of existing traffic signals. For a comparison with conventional approaches, see what advantage AI provides over traditional traffic management.
Public transport: forecasting demand and predicting bus delays
AI helps public transport mainly through forecasting: predicting passenger demand per station so frequency and stop spacing can match it, and predicting bus delays so riders see realistic arrival times. Schedule optimisation from predicted demand follows from that forecast.
On delays, Google’s bus model combines real-time car traffic forecasts with route and stop data, and it has to translate car speeds into bus speeds differently on different streets. One finding deserves attention from anyone building this: in Google’s experiments, adding official agency schedules did not significantly improve the forecasts. The model also has to generalise to cities with no real-time feed, because many agencies cannot provide one.
On demand, a memory-augmented multi-task LSTM tested on Greater Sydney data (bus, train, light rail and ferry, 6.37 million users) improved MAE by 7.75% and RMSE by 14.98% over a multi-task LSTM baseline, by transferring patterns from the dense bus network to sparser modes. Its ferry forecasts were weaker because weather affects ferry demand, and irregular days were hard to predict without additional inputs. That is the general lesson in one case: the model is only as good as the signals it is given about what moves demand. Travel operators face related passenger-demand forecasting questions, covered on our page on AI engineering for travel and hospitality.
Fleets: predictive maintenance and routing
Predictive maintenance uses mileage patterns, diagnostic trouble codes, sensor readings and engine trends to flag a developing fault in a specific vehicle before it fails, rather than servicing on a calendar. Telematics collects parameters such as engine temperature, oil pressure, battery voltage and brake wear continuously, software compares them with patterns that historically precede a breakdown, and the alert reaches the maintenance team with enough lead time to schedule the repair.
Two practical points decide whether a programme works. Mixed-age fleets have uneven diagnostic data, so older vehicles may need retrofitted telematics before they can be monitored. Accuracy has to be tracked after launch: how many predicted failures were confirmed, how much downtime was avoided, and which breakdowns the system missed, with alert thresholds adjusted once months of real data exist.
Routing is the other fleet workload: AI-driven navigation systems draw on GPS traces, traffic reports and weather to order stops and adjust routes as conditions change. An NMFTA analysis cites a Gartner study from February 2024 finding that as many as 45% of transport and supply chain companies use AI for demand prediction.
How AI driving works in an autonomous vehicle
Waymo’s autonomous vehicle runs a perception, prediction and planning stack in real time. Waymo describes its perception system identifying pedestrians, cyclists, vehicles, construction and signal states from its sensor suite; its prediction stage forecasting the possible paths of other road users; and its planner choosing trajectory, speed, lane and steering from maps, detected objects and those predictions. Localisation matches detailed custom maps against live sensor data rather than relying on GPS alone. An onboard computer with server-grade CPUs and GPUs runs this, with a secondary computer designed to bring the vehicle to a safe stop if the primary fails.
Research is also pushing end-to-end systems that map raw sensor input directly to motion plans, gaining joint optimisation of perception and planning at the cost of open problems in interpretability, causal confusion and robustness. Our article on AI for autonomous vehicles goes further into the vehicle side.
Over 56.7 million rider-only miles to January 2025, Waymo’s crash rates were statistically significantly lower than human benchmarks for any-injury-reported, airbag-deployment and suspected-serious-injury crashes, with a 96% reduction in injury-reported vehicle-to-vehicle intersection crashes and no crash type showing a statistically significant disbenefit. That is one operator in its own service areas, which is a boundary on how far the result transfers. Rare events remain hard to capture: one study cited in a 2024 survey puts the ratio of crash to non-crash data at 1 to 11,000. Waymo reports more than 20 billion simulated miles alongside over 200 million real ones.
What role does GPU acceleration play in transport AI?
Real-time transport AI leans heavily on GPU acceleration. A 2023 survey of edge video analytics gives less than 30 ms end-to-end latency as the requirement for applications like autonomous driving, and notes that large-scale real-time tracking of road users in traffic monitoring relies heavily on GPUs. Deep learning has outgrown ordinary CPUs; mainstream compute is GPU and TPU.
One central hardware question is where inference runs: at the edge or in the cloud. A 2026 city-scale system built on Bengaluru’s Safe City cameras runs detection and tracking on Nvidia Jetson devices at the edge, sending compact vehicle-flow summaries instead of raw video, and reports stable ingestion up to 2000 FPS on Jetson Orins with graph neural network forecasts scaling to 1000 streams. Its forecast RMSE rose modestly from 20 at a one-minute horizon to about 23 at four minutes. Bellevue, Washington, ran a similar edge pattern on Azure Data Box Edge with the cloud as backup. FPGAs and Google’s Edge TPU fill parts of the same niche for detection workloads.
The engineering tension is model size against the latency budget: heavy networks meet accuracy targets and miss latency, light ones do the reverse. Practical answers include cascades, where a small model handles easy frames and a heavier one runs only on low-confidence results, and choosing lighter architectures for small devices; one Scientific Reports study found GPU training cut LSTM preparation time by about 400 times but recommended GRU over LSTM for edge deployment because of LSTM’s computational needs. Swapping models at runtime has its own cost, since loading a new DNN into memory can take up to a second. The same edge-versus-cloud trade-off shows up at household scale in smart home devices.
Where does generative AI fit into traffic and transportation?
Generative AI’s advantages in transport lie in data generation, reconstruction and uncertainty modelling. A 2024 survey of generative AI for intelligent transportation systems groups its uses into perception, prediction, simulation and decision-making, and they include filling gaps: imputing missing sensor data, estimating traffic where no sensors exist, and generating synthetic driving scenes and crash data. For autonomous driving, generated images and videos cover weather, lighting and scenarios that are expensive or dangerous to record.
Two limits matter for anyone planning to rely on it. Generated traffic data can look realistic while lacking spatial or temporal consistency, and generators tend to reproduce common patterns while neglecting the rare cases the data was meant to supply. LLM-based driving and traffic frameworks report high accuracy but struggle with latency and resource demands, and still need more robust, responsive decisions in unfamiliar scenarios. In the Wisconsin DOT survey, generative AI scored the highest potential benefit of the six domains (4.32) yet sat among those recognised for long-term potential with higher uncertainty. Our reading is that synthetic data for training and testing is the near-term use, and generative models in the real-time control path are some way off.
What are the main challenges of deploying AI in transport systems?
The main challenges are data quality, real-time performance, security, privacy and legal compliance, interpretability and workforce skills.
- Data quality. Traffic data is often sparse, noisy or missing because of sensor failures, and supervised models need large labelled sets. California’s PeMS data, used in many studies, regularly contains stations that fail to report reliably. Vision data is valuable but labelling it frame by frame is slow and needs domain expertise.
- Latency and compute. Real-time perception has to cope with imperfect inputs while staying fast enough to act, which forces a trade between model complexity and response time.
- Security. V2I and V2V links are attack surfaces that need mutual authentication, and SAE J3061, built on ISO 26262, distinguishes safety (faults) from cybersecurity (purposeful attack). Backdoor attacks through poisoned training data could make a monitoring system misclassify vehicles or an autonomous vehicle make wrong decisions.
- Privacy. In January 2026 the FTC finalised an order requiring GM, for 20 years, to obtain affirmative express consent before collecting, using or sharing connected vehicle data, after allegations it sold geolocation and driving behaviour data without consent. Predictive maintenance and autonomous driving systems collect data on driver habits and operations, so the same exposure applies to fleets.
- Legal compliance and liability. A November 2023 DigitalOcean study found almost 30% of surveyed companies named ethics or legal compliance as main concerns preventing AI adoption, and US state AI rules differ widely. Liability after an autonomous vehicle crash remains a major barrier.
- Crash reporting. NHTSA’s Standing General Order requires named entities to report a crash if an automated driving system was in use within 30 seconds of it and the crash met damage or injury thresholds.
- Interpretability. Deep traffic models are generally more accurate than classical ones but offer no explanation of why, which makes reliability hard to verify.
- Skills and cost. 73.3% of Wisconsin DOT survey respondents identified technical skills as a gap, and half named data management.
Autonomous vehicles stand apart in that survey: they were the only domain where perceived risk (4.15) exceeded perceived benefit (3.95). The report’s recommendation follows from that: scale mature applications such as asset management and safety, and pilot autonomous vehicles and generative AI in phases.
Separating a deployable system from a promising pilot
Surtrac’s operators found limitations only during continued operation: detection errors, a bridge closure that reshaped flows, and buses blocking lanes. The useful question for any transport AI proposal is therefore not how big its headline gain is, but whether that gain was measured on live traffic, what happens to it when a sensor drops out, and whether the inference hardware still meets its latency budget at the camera count the city actually plans to run.
