How AI predicts when your ride will arrive

Your car is getting closer, yet its arrival time moves later. That can be a reasonable update. Uber distinguishes the estimate for a driver reaching your pickup from the estimate for reaching your destination once travelling. Both concern a changing journey: traffic and roadworks can alter what lies ahead.[1]

An arrival estimate has to connect observations made now with events that have not happened yet. How long will a road take when this vehicle reaches it? Which route is usable? Several models and calculations can contribute to the answer.[3]

First, the vehicle has to be on the right road

A position reading does not arrive labelled with the correct road segment. In its 2019 CatchME account, Uber described matching noisy GPS observations to plausible road sequences using a probabilistic model. Buildings and other sources of location error complicate this step. A point appearing beside a street does not necessarily mean the driver left it.[2]

There is a second possibility: the map itself is wrong. CatchME looked for repeated disagreement between suggested routes and observed travel, helping operators investigate incorrect road restrictions. One odd trace could be noise; consistent traces can justify checking the map. A precise traffic forecast cannot make an unusable route usable.[2]

The forecast must look ahead along the journey

The Google Maps research published in 2021 represented connected road segments as a graph. Its model learned travel times for groups of segments at different future horizons. A proposed journey could then query later sections for later times, rather than assume that every road would have its present traffic when the vehicle arrived.[3]

Connections matter because traffic on one road affects its neighbours. The researchers’ approach combined recent conditions with historical patterns and shared a model across differently shaped road groups. That lets the system learn relationships across the network, including how congestion around a junction can affect the route being timed. These are documented research mechanisms, not a claim that Google’s entire current system is unchanged.[3]

Uber’s May 2026 DeepETT account describes a newer upstream forecaster. It refreshes road-segment travel-time predictions every five minutes, reaching up to three hours ahead. A transformer combines historical and recent signals with information about nearby roads and wider geographic regions. Its numerical inputs describe traffic, rather than words in a conversation. Routing uses these forecasts to calculate a journey.[4]

The trip gets a correction of its own

Uber’s 2022 DeeprETA paper explains the next distinction: a route calculation and a real trip need not agree. Drivers can take a different route, and errors can accumulate across segments. The model learns a correction to the route-based estimate using information such as origin, destination, time and trip type. The correction is added to the initial estimate.[5]

This separates two useful questions: what the road network appears to permit, and what completed trips teach about the remaining error. The combination preserves a structured routing system while learning recurring differences between its calculation and observed journeys.[5]

A better component can upset the final answer

DeepETT exposed an instructive problem. Uber reports that improved traffic predictions initially reduced final ETA accuracy, even after retraining the downstream model. The components’ calibration could drift between training cycles and vary by place and time. Recent observed-versus-predicted travel times supported real-time corrections. Improving a road-segment score therefore did not automatically improve the number shown to the traveller.[4]

Read the arrival time as a fresh prediction

In this invented example, three minutes pass while the remaining estimate falls by only one minute. The predicted arrival consequently moves two minutes later. At the next update, it moves earlier again. These are consistent calculations. A countdown to a fixed event and a forecast revised using new information are different objects.[1]

The target can also differ. DeeprETA describes training objectives that penalise early and late errors differently. Estimating a typical arrival and choosing a more conservative planning time therefore need not produce the same answer from the same evidence.[5]

For your own plans, decide which event matters: meeting the driver, reaching the drop-off, or being seated at an appointment after a walk. Treat a revised estimate as new information about that journey. Its usefulness comes from how well the prediction follows the changing trip, with the destination and timing question kept clear.

Report an error or suggest a correction