AI in Transportation: Driving Is Not Route Optimization

As of July 30, 2026, automated driving and route optimization share maps and traffic data but solve very different decisions. A vehicle must react within milliseconds to a pedestrian; a fleet spends minutes assigning stops under capacity and time windows. Machine learning may estimate objects, trajectories, travel times, or demand. Planning and control then turn estimates into actions. Separating estimator from optimizer prevents one network from being credited with “autonomy.”

Two scales, two contracts

In driving, the final output is a trajectory and physical commands within an immediate environment. Errors can cause harm before a person intervenes. In logistics, the output is a plan of visits, vehicles, and times; it can be reviewed, reoptimized, and compared by cost.

For every project, write the decision, horizon, frequency, constraints, owner, and fallback. If frequency is 20 Hz, latency belongs to accuracy. If routes are rebuilt hourly, plan stability and the cost of changing commitments matter alongside distance.

The operational domain limits driving

“Autonomous vehicle” does not say where it can operate. NHTSA distinguishes driver assistance and automation and stresses that available technologies have different capabilities and responsibilities. A system should declare roads, speeds, weather, lighting, regions, and states where it functions: its operational design domain.

Outside that domain, it must detect the condition and reach a minimal-risk state. A demonstration on a dry motorway does not authorize unmarked streets in snow. Coverage is evaluated against the declared domain, and every expansion opens another test program.

Sensors: redundancy is not repetition

Cameras supply texture and color; radar measures range and relative speed; lidar supplies geometry; GNSS and inertial sensors locate and estimate motion. Sensor fusion requires spatial and temporal calibration. Two sensors may share a failure because of dirt, weather, or bad synchronization.

Useful redundancy detects disagreement and preserves a minimum function. Tests inject failures: blocked camera, ghost return, degraded position. The system should raise uncertainty, slow down, or hand over rather than silently average its way to false certainty.

Perception: from pixels to objects and free space

Detection, segmentation, and depth produce representations of the environment. Reference datasets help measure components. KITTI collected camera, laser, and GPS data for mobile-vision tasks; nuScenes expanded sensor combinations and annotations in urban scenes.

High mAP is not driving safety. Results should isolate distant pedestrians, small objects, night, occlusion, and unknown classes. An out-of-distribution input should raise uncertainty or trigger a conservative policy. The open world contains objects absent from the ontology.

Prediction: other actors make decisions too

After detection, the system estimates how pedestrians, cyclists, and vehicles may move. One trajectory confuses uncertainty with certainty; several plausible futures and calibrated probabilities are needed. The Waymo Open Motion Dataset was built to study interactive prediction through scenarios and maps.

Average geometric error can reward a trajectory passing between two possible futures while representing neither. Evaluation needs mode coverage, collisions, calibration, and performance by maneuver. Prediction feeds planning; if it omits a dangerous possibility, planning cannot compensate.

Planning and control: feasibility before comfort

A planner chooses a trajectory respecting roads, dynamics, margins, and rules. A controller turns that trajectory into steering, acceleration, and braking and corrects error. An end-to-end model can learn part or all of the chain, as End to End Learning for Self-Driving Cars demonstrated in a bounded setting; that does not remove the need to measure interfaces, limits, and failures.

Independent constraints may check speed, acceleration, distance, and reachable space. Fallback must be rehearsed at real latency. Comfort—jerk, oscillation, harsh braking—is measured with progress and safety. Arriving without a collision is insufficient if behavior is unstable.

Rare events require simulation and reconstruction

Many ordinary miles provide little evidence about an infrequent event. Scenarios should come from crashes, near misses, systematic variations, and adversarial combinations. CARLA provides an open simulator with traffic and sensors for urban-driving research.

Simulation alone does not certify safety because pedestrians, friction, and sensors remain models. It connects to track repetition, shadow mode, and limited deployment. Every incident requires synchronized logs for sensors, perception, prediction, plan, control, intervention, and version so the chain can be reconstructed.

Route optimization begins as a combinatorial problem

In basic routing, an algorithm seeks a low-cost path through a network. A fleet adds vehicles, depots, capacity, shifts, time windows, pickups, and deliveries. The OR-Tools Vehicle Routing Problem shows how these constraints make the task different from predicting traffic.

The optimizer receives costs and constraints and returns a plan. It may be exact for small instances or heuristic for large ones. Baselines include current routes, nearest neighbor, or another solver under equal time. Report distance, duration, vehicles, violations, compute time, and stability.

Where learning enters routing

A model can estimate travel time from segment, hour, weather, and history; forecast orders; predict cancellations; or estimate service duration. DeepTTE studied end-to-end travel-time estimation from trajectories. Those outputs become optimizer parameters; they are not routes by themselves.

Evaluation must cross the interface. A small ETA gain may reduce late arrivals or change no route. Compare four combinations: baseline predictor plus optimizer; new model plus the same optimizer; the same costs plus a new optimizer; and the current system. That attributes the gain.

Uncertainty and replanning

A mean time hides congestion tails. Quantiles or scenarios allow margin and estimation of on-time probability. A plan may optimize expected cost, a reasonable worst case, or a service level. That preference is operational rather than a model property.

New information triggers reoptimization, but continuously changing drivers and stop order may be worse than accepting a short delay. The objective should penalize plan disruption. Every recommendation still obeys legal constraints, rest, capacity, and access.

Clustering assigns neither meaning nor route

An unsupervised algorithm may group stops with similar patterns or vehicles with similar cycles. As the scikit-learn clustering documentation makes clear, clustering finds structure under selected data and distance. It does not independently decide that a group is “priority,” nor does it produce a feasible route.

A person or rule interprets clusters and the optimizer applies constraints. Scaling, distance, and cluster count change the result. Geography must be checked: two points close in a straight line may be separated by a river or road restriction.

Fleet data changes the system too

Incomplete GPS, switched-off phones, new drivers, and low-coverage areas create bias. Measuring only instrumented vehicles may portray a partial city. Location is also personal and workplace data; retention, access, and purpose need definition.

An optimization may cut mileage and worsen workload or safety by concentrating impossible routes. Results should be broken down by zone, shift, and vehicle type, with an exception mechanism. The objective function materializes human priorities.

The test that keeps both problems separate

Draw two chains. Driving: sensors→perception→prediction→plan→control→fallback. Routing: demand/ETA→costs and constraints→solver→plan→replanning. For every arrow, record uncertainty, latency, metric, and owner. If learning disappears from one box, ask which deterministic algorithm remains; a system often still does.

The durable skill is distinguishing prediction from decision. Knowing where traffic exists neither brakes a car nor assigns a fleet. Evidence arrives when each estimator is tested under its distribution, every optimizer respects constraints, and the combined system has a safe exit when the world does not resemble training.

A final cross-layer test deliberately degrades one estimate at a time. Add delay to ETA forecasts and inspect missed windows; hide or perturb an object and inspect the driving fallback. The plan should degrade in a controlled way rather than fail silently. This experiment connects component metrics to operational consequences and reveals whether uncertainty actually travels through the system.