ЭвоКарго: почему локализация автономного транспорта — это борьба с противоречивыми данными

EvoCargo's localization and mapping team has published the second part of a series about how autonomous vehicles understand their location. If the first part was about navigation methods, this one is about data: why its processing is more complex than it seems, and how the system works despite constant contradictions.

More sensors — does not mean more accurate

A common misconception: the more sensors, the higher the accuracy. In theory, it makes sense. In practice, each sensor lives in its own "reality" — with its own physical limitations, characteristic errors, and specific degradation scenarios. Here's what the localization system faces in real conditions:

• LiDAR — loses accuracy in snow, fog, and rain; creates glints on reflective surfaces
• GPS/GNSS — disappears under overpasses, in tunnels, and near tall buildings
• Cameras — dependent on lighting, struggle with oncoming headlights and lens flare
• IMU (inertial measurement unit) — does not require external signals, but accumulates integration error over time
• Odometry — accurate over short distances, but wheel slip and uneven surfaces introduce systematic errors

When several such sources work simultaneously, the system receives not just noisy data — it receives contradictory statements about its own position in space.

Algorithm as arbiter of contradictions

The authors describe the algorithm's state through the image of Travolta from "Pulp Fiction" — that very meme where the character looks around in bewilderment. This is exactly how the system "feels" when multiple sensors simultaneously say different things: where the vehicle is, where it's moving, and at what speed. The solution is not to choose the "correct" sensor and ignore the "wrong" one, but to combine all sources in a weighted manner, accounting for their current reliability. This approach is called sensor fusion. The word "current" is key here: sensor reliability changes in real time depending on environmental conditions.

"Localization is not about measurements.

It's about interpreting contradictory data under conditions of uncertainty hundreds of times per second," — EvoCargo team.

What happens inside the system

The article provides a detailed breakdown of the architecture that allows EvoCargo vehicles to operate in real warehouse and terminal conditions. The system solves several tasks simultaneously. First, it assesses the credibility of each sensor during movement — not based on datasheet specifications, but on how the data behaves at the current moment.

If LiDAR starts producing atypical "noise," this is a signal of degradation before formal failure. Second, it combines data with different timestamps and sampling rates: GPS updates once per second, IMU hundreds of times. Reconciling them to a single point in time without accumulated error is a non-trivial mathematical task.

Third, it interpolates the vehicle's position between measurements so that the control system always knows the current coordinates, not those recorded half a second ago.

What this means

Reliable localization is not a race for the number of sensors. It's the ability to build mathematically robust algorithms that work precisely when data is unreliable. For industrial autonomy, this is perhaps the most important engineering skill — and it's exactly what determines the difference between a laboratory prototype and a system that works every day in any weather.