There’s no such thing as mapless automated driving

With AI taking on a larger role in automated driving (AD), expectations of what autonomous vehicles should be capable of are rising. Vehicles are expected to go beyond reacting to what's ahead, and even predict what will happen next. In this growing reliance on neural networks, maps have increasingly been labeled as optional, or even obsolete. But as with every new tech, there's a path to evolution until things work perfectly. The modern world of safe automated driving still depends on a contextual awareness and grounded understanding of what lies ahead — how the road is shaped, how traffic typically behaves and what situations are likely to unfold beyond the range of onboard sensors.

AI automation and looking ahead with maps 

As AD technology develops, the industry is beginning to see a division between those that believe AD needs a map and those that don't. Today’s AD systems rely heavily on sensors, especially in the perception layer — the part of the stack that turns raw sensor data into an understanding of the vehicle’s environment. To be certain that a vehicle acts safely, however, the perception stack relies on more than just sensors — it also incorporates data from maps.  

In most advanced driving systems today, digital maps provide structure and context that sensors alone cannot always capture, or might fail to, especially when obscured by unfavorable weather conditions like rain or snow. They tell the vehicle the lane it should be in, what the speed limit is, how intersections are shaped, where road edges are and what to expect beyond the range of the cameras.  

“

In the perception stack, if LiDAR gives you the immediate view of the ‘here’ and ‘now’, maps really help you understand what’s coming next.

Andrew Hart

CEO, SBD Automotive

Hart believes that supporting the vehicle’s understanding of the world with maps is key to creating more trust in the vehicle, which remains a barrier in the advancement of AD technology.

At the same time, thanks to AI, a new generation of automated driving software is emerging — end-to-end (E2E) AI models. Instead of splitting autonomy into separate modules for perception, prediction and planning, E2E AI models process sensor data within a single large neural network, allowing vehicles to make more dynamic decisions. Unlike traditional modular AD systems, these AI-driven models have a simpler architecture and can adapt and learn from their own experiences of the world. Similar to large language models (LLMs), these networks ingest sensor data and output driving decisions. And while this sounds impressive, the inexplicable, ‘black-box’ nature of this tech has raised safety concerns. The highly automated and data-driven nature of these models makes it difficult to understand its internal decision logic, which means it can be extremely difficult to understand why a vehicle acted in a certain way, especially if it behaved unsafely.  

Modern AI systems can also misclassify objects, produce false positives or create scenarios that may exist in real life — sometimes described as ‘hallucinations’ — and in the case of autonomous vehicles, mislabeling objects or hallucinating a pedestrian could have catastrophic consequences. 

Can autonomous vehicles go mapless? 

“Just like we don’t want LLMs to make things up, we don’t want these AD models predicting fantasyland ahead. We want them to predict reality. And the map is what grounds reality,” he says. 
Maps allow the vehicle to communicate its 'thinking' to the driver and passengers, helping maintain trust in the vehicle.Harrell adds that maps remain essential not just for how the vehicle perceives the world, but also for how it communicates that understanding back to the human inside. Even if an E2E AI model is making the predictions, the system still needs a way to show the driver what it ‘sees’, what it plans to do next and how it interprets the road ahead.  
“How do you display that? Through the map onto the screen,” he explains. “The map is the interface between the model and the operator, so it still has to exist.” 

To support both the vehicle’s intelligence and the driver’s trust, automakers need a map that is detailed enough for machine reasoning and clear enough for human interpretation — a map that is both technically robust and constantly up to date. 

Automating mapmaking to accelerate vehicular automation 

Instead of only relying on survey vehicles, sensors and manual editing, TomTom’s Orbis Maps for Automation uses a combination of foundational models and in-house AI to automatically process multiple data sources and generate a highly detailed, 3d map of every road, down to lane-level precision.  

The latest example of this approach is CARIAD, the automotive software company of the Volkswagen Group, which has selected Orbis Maps for Automation as a core component of its AD systems. By integrating Orbis map data into its AD stack, CARIAD will strengthen how its vehicles understand and anticipate their surroundings, using real-world spatial context to support functions ranging from Intelligent Speed Assistance to hands-free driving. In doing so, the map acts as a stabilizing layer — helping automated systems navigate complex driving scenarios in a predictable, human-like manner. 

The map in ‘mapless’