Positioning and Vehicle Motion

An automated vehicle must estimate not only what surrounds it, but where it is and how it is moving. Positioning is therefore a perception problem with its own uncertainty, failure modes and redundancy requirements.

The vehicle state

The navigation and driving stack may need estimates of position, heading, speed, acceleration, yaw rate and wheel motion. No single sensor provides all of these reliably in every environment.

GNSS supplies a global position and time reference, but buildings, terrain, tunnels, interference and reflected signals can degrade it.
Inertial measurement units measure acceleration and rotation at high frequency. They bridge short gaps, but small biases accumulate into drift.
Wheel-speed and steering sensors support odometry and motion estimation. Slip, tire changes and uneven surfaces introduce error.
Cameras, radar and lidar can estimate motion relative to landmarks or a map.
High-definition maps can add lane geometry, road boundaries and other prior information, but they can be stale or locally wrong.

The result is normally a fused estimate with a confidence or protection level, not one exact coordinate.

A consumer navigation receiver can select a route even when its position is uncertain by several metres. A driving system that plans a trajectory within a lane may need much stronger lateral and heading accuracy, plus integrity: it must know when the estimate cannot be trusted.

Point-to-point navigation also does not define the automation level. A Level 2 system may follow a navigation route and execute steering and speed actions while the driver remains responsible for continuous supervision and fallback. The distinction is explained in SAE Level 2 versus Level 3.

Relative and absolute localization

Absolute localization relates the vehicle to a global frame, usually through GNSS and a map. Relative localization estimates movement from a prior pose using inertial sensors, wheel odometry or changes in perceived features.

Modern systems combine both. When GNSS is weak, inertial and scene-based estimates can bridge the gap. When a distinctive mapped landmark becomes visible, it can reduce accumulated drift. The software must also reject a bad correction rather than forcing the estimate onto the wrong road or lane.

Maps are sensors with a long delay

A map is not a live measurement. It is stored prior knowledge collected earlier, distributed later and localized against current observations. Construction, temporary lane markings, moved barriers and changed speed limits can invalidate it.

The correct relationship is “map plus live perception,” not “map instead of perception.” The system needs rules for conflict: live road closure signs and barriers should not be ignored merely because the map expects an open lane.

Integrity and degraded operation

Positioning safety depends on more than average error. Engineers also need to understand rare large errors, spoofing or jamming, multipath, sensor bias and correlated failures. A system should monitor whether the available measurements agree and whether the resulting confidence remains sufficient for the active function.

If localization quality falls below the requirement, a Level 2 feature may alert the driver and disengage. A Level 3 or Level 4 system needs behavior consistent with its fallback responsibility and operational design domain, which can include reducing speed or reaching a minimum-risk condition.

Mercedes-Benz describes DRIVE PILOT as combining high-precision positioning, an HD map and onboard sensing. Its automated-driving safety material also discusses GNSS, odometry and landmark-based localization. See Mercedes-Benz: DRIVE PILOT technology and Safety First for Automated Driving. These are examples of one system design, not universal prescriptions.

What to ask about a positioning claim

Is the stated accuracy typical, worst-case or a laboratory result?
Does it describe global position, lane assignment, heading or integrity?
Which correction service, map and road environment were used?
What happens in tunnels, urban canyons or changed road layouts?
How quickly is a bad estimate detected?
Which function remains available after GNSS, map or landmark localization is lost?

The positioning stack belongs inside the complete sensor-fusion architecture and safety argument. A centimetre-level headline is not useful if it is unavailable, undetected when wrong, or unrelated to the function's actual requirement.