Michael Harrell is the SVP of Engineering Maps at TomTom.
getty
AI’s growing role in vehicles has fueled claims that maps may soon become obsolete. With advances in sensors, end-to-end (E2E) AI and predictive models, a narrative has emerged in autonomous vehicle (AV) development that labels maps as optional, arguing that vehicles will largely infer the world from what they see in the moment.
But seeing the road isn’t the same as knowing the road. Sensors provide a snapshot of the present, but safe automated driving also depends on anticipation, memory and verified context. A vehicle that relies only on real-time perception drives like a tourist, discovering every curve and restriction as it goes. One supported by map intelligence behaves more like a local, anticipating what lies ahead.
This is the crux of the “mapless” debate: Maps, including high-definition (HD) or lane-model types, back most autonomous use cases today, whether for training, evaluation or real-time inference. Even systems described as “mapless” are often developed or tested using mapped representations of the road; in nearly every case, a map is still part of the system. They’re invisible infrastructure, helping vehicles understand the road beyond immediate perception.
The debate, then, isn’t really whether maps exist in the stack, but how much detail is required at each stage to deliver the necessary performance, safety and scalability.
From Perception To Reliable Performance
E2E AI processes combined sensor data to make driving decisions and allows vehicles to perceive and decide as a more cohesive system. Instead of separate systems for lane detection, object recognition and path planning, a single model interprets all inputs at once, such as how vehicles interact, how pedestrians behave and how road geometry shapes movement. This provides new capabilities in dense urban settings where modular systems have struggled, such as autonomous overtaking.
Although these advances are significant, Level 3 automation in consumer vehicles remains limited, and for safety-critical systems, every incremental improvement becomes considerably harder to achieve. Map intelligence can provide advanced knowledge of road geometry, lane structure, speed changes and restrictions outside immediate sensor range. This doesn’t replace perception—it gives perception a continuous reference layer against which to plan, compare and validate.
MIT research illustrates the value of that combination. Researchers developed an autonomous navigation system that used camera data alongside a simple GPS-like map to drive through previously unseen environments. The system continually compared visual observations with map information and detected mismatches, helping it recognize when its position, sensors or map inputs might be wrong, ultimately helping AI systems check their own understanding of the world.
Why Maps Still Matter In An AI-First World
Sensors can only perceive what lies within their physical range. They can’t see around corners, beyond large vehicles or through temporary obstructions. To drive safely, AI systems must understand what comes next—upcoming turns, hidden hazards, shifting lane patterns or temporary restrictions—and self-learning systems remain imperfect enough that inaccurate outputs carry real consequences in a safety-critical application like driving.
An IEEE Computer Society study notes that HD maps are used in AV perception stacks because they encode road-level features with centimeter-level accuracy. The challenge is ensuring that map data is current, interoperable and suited to its intended function.
The economics and speed of map creation have also changed. Today, fleet cameras, onboard sensors, satellites and crowdsourced observations can help create and refresh maps at far greater scale (paywall). Texas A&M researchers, for instance, are developing a framework for crowdsourced 3D map generation and visual localization using information collected by vehicles already on the road.
Continuous data collection and automated change detection make it easier to identify discrepancies, validate updates and distribute information faster. These capabilities don’t guarantee “four or five nines” of reliability, but they can help reduce uncertainty and strengthen data for training, evaluation and operation.
The “Mapless” Debate Is Really A Map-Detail Debate
The level of map detail and intelligence needed at each stage of automation should reflect the driving task. During training and evaluation, automated systems benefit from rich representations that help them learn and test relationships among lanes, signs, speeds, traffic patterns and driver behavior. At runtime, a vehicle may need only selected layers for a particular operational design domain.
A vehicle that brakes late or changes lanes for no clear reason can feel unsafe even when it avoids a collision. As drivers become less active, system behavior must become more explainable, and a reliable world model gives automated systems a consistent reference point for planning and validation.
The Road Ahead
As vehicles progress from assistance to higher levels of automation, AI will continue to redefine what machines can perceive and predict. But perception alone isn’t enough. Automated systems must be grounded in accurate, continuously updated representations of the physical world.
This path is already visible, with higher automation appearing in limited use cases such as driverless taxis and controlled environments like mining and ports. Scaling to broader everyday driving requires systems that can handle constantly changing road realities.
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?
