Some work in the ULTIMO project is dedicated to establishing a standard for HD maps for autonomous transport, The goal is to reduce Automated Driving System (ADS) deployment costs and enable mixed-manufacturer Automated Vehicle (AV) fleets, thereby reducing potential customer lock-in caused by high mapping costs.
After three and a half years on the job, we know that a common standard for HD maps for autonomous driving is beyond reach. The problem is both technical and commercial. In addition, there is a fundamental industry split and a great deal of confusion that prevents any common solution.
What is going on behind the scenes that we did not see at the beginning? The industry is divided into two parties. Providers such as Waymo, Mobileye, WeRide and also European manufacturers such as NAVYA, AuveTech, EasyMile and others rely on preliminary maps. This means they map their Operational Design Domain (ODD), which defines the conditions under which automated driving functions may be used, before operations begin. Other players such as Tesla and additional technology driven vehicle developers are pushing for a completely mapless future that mostly relies on camera based localisation.
However, the situation is more complex than simply maps versus no maps. The confusion is accentuated by three major reasons.
First, the term HD maps is being used by both Advanced Driver-Assistance Systems (ADAS) and ADS manufacturers such as Waymo, Mobileye, NAVYA, AuveTech and other European manufacturers, and by map providers such as TomTom and national cadaster services, as if it refers to the same thing. TomTom advertises HD maps for automated driving, yet the reality is that the HD maps required for assisted driving systems are fundamentally different from those required for fully automated deployment in a defined Operational Design Domain. In practice, we are speaking about different products with different purposes, accuracy requirements, update mechanisms and liability implications. Treating them as one and the same creates conceptual confusion from the outset. Furthermore, each ADS manufacturer optimizes the HD map for its specific ADS by merging sensor data used for map creation into models that are tightly coupled with the ADS software.
The second reason is that vehicle manufacturers, with Tesla leading, blur the lines in their marketing by presenting Level 2 and Level 3 vehicles as autonomous vehicles, although they are assisted driving vehicles. This distorts the debate. Level 4 vehicles operate autonomously within a known and predefined area. Level 5 vehicles are intended to operate anywhere, in unknown environments. When assisted driving systems are presented as autonomy, the discussion about maps becomes confused because the operational assumptions are entirely different.
A third reason is the misunderstanding of the target vehicle use. While the automobile industry aims to sell private cars that roam the whole planet, public transportation needs vehicles that are confined to a well-defined deployment area. This is the key difference between Level 4 and Level 5 automated vehicles in practice. Level 4 vehicles are deployed in a known area that must be scanned in advance. Level 5 vehicles are meant to go anywhere without prior mapping. The interest of the automobile industry is clearly to provide vehicles that can go anywhere and are not bounded by pre scanned areas. For this reason, mapless solutions are promoted, although they are far more complex to implement.
Public transportation operates under different constraints and objectives. It provides deployment in predefined areas and must integrate into city maps soft annotations that are specific to public transport services. In addition, public transportation requires fleet orchestration and demand responsive transport systems to manage dynamic rerouting and operational optimisation with shared use of the vehicle. These requirements do not apply to private vehicles in the same way, as their operation typically involves simple point-to-point transportation. Fleet orchestrators require maps, but not necessarily vehicle grade HD maps. At the same time, the vehicle HD map must be coordinated live and dynamically with the fleet orchestrator in order to achieve optimal operation.
After three and a half years of work, it has become clear that the barrier to scaling is not only technical complexity but also business models. ADS manufacturers relying on preliminary map providers consider their data and the optimizations made to enhance ADS performance as a competitive advantage, while treating interoperability as a commercial risk rather than a technical goal. As long as ADS manufacturers consider their HD maps a competitive advantage instead of a shared utility, standardisation will remain extremely difficult.
It is clear today that full-stack sharing of HD maps for ADS will not be possible in the near future, as ADS manufacturers will continue improving their ADS and corresponding HD maps to provide a more competitive offering. At best, the higher levels of HD maps and user annotations can be standardized, partially reducing deployment costs. If we want full interoperability of ADS HD maps, we need more than new standards. We need clarity about definitions, honesty about levels of automation and, above all, a decision about whether digital infrastructure for autonomous transport should be treated as a private asset or as a public utility. A truly interoperable approach can only be considered once this question has been addressed.
The photo above was taken at the GAMMS (European Horizon 2020 project “Galileo/GNSS-based Autonomous Mobile Mapping System”) final event.