The Future of Mapping: Maps Read by Cars, Drones and Robots

For most of their history, maps were drawn for people. A person looks at a paper map or a screen, sees a thick yellow line, and understands "main road". Today the heaviest map readers are machines: the speed assistant in a new car, a robotaxi planning a lane change, a sidewalk robot looking for a kerb ramp, a drone checking whether it may fly over a hospital, and an AI agent answering "which pharmacy is open now".

Each of those machines reads different parts of the map, and none of them can look at a yellow line. They need attributes: how many lanes, what speed limit, which side has a sidewalk, how high the kerb is, how tall the building is. We measured how fast that detail is arriving in open map data. Between January 2016 and September 2026, OpenStreetMap objects in Europe with a lane count grew 3.7 times, sidewalk tags 8.8 times and kerb tags 83 times, while the number of road objects grew 2.5 times. The map is getting deeper faster than it is getting bigger.

This article is the start of a series about the machines and businesses that read maps. It explains what "machine-readable" means, what each kind of machine needs, how those maps are made, and where European rules come in.

What this series covers

Over the coming months we publish one article per machine or use case, each with its own measurement:

• the levels of self-driving cars and what each level asks of the map
• GPS jamming and spoofing, and how vehicles cope without satellite positions
• telematics and fleet data
• digital twins of cities
• drone delivery, drone no-fly zones and air taxis in Europe
• route planning for electric vehicles
• lidar mapping and RTK positioning
• turn-by-turn navigation and AR navigation
• self-driving trucks and delivery robots
• location data privacy, V2X and smart city data
• how maps stay up to date, indoor positioning and micromobility

Who reads the map now

Each reader needs a different slice of the same map. The table sums up what each one reads and how fresh it has to be.

Two points come out of the table. First, machines need attributes more than geometry. A robot can see a sidewalk with its cameras, but it plans its route before it leaves the depot, and for that it needs to know which streets have one. Second, every reader cares about age. A speed limit from 2019 is a liability in a car that warns the driver about it.

The articles on HD maps for autonomous driving and intelligent speed assistance go deeper on cars. The robotaxi cities tracker shows where driverless vehicles already carry passengers.

What makes a map machine-readable

A map is machine-readable when software can answer a question from it without a human interpreting a picture. In practice that takes five properties.

• Vector data. Roads are lines, buildings are polygons, a fire hydrant is a point. A raster image of a map tells software where the pixels are; a vector map tells it what the objects are.
• Attributes on every object. maxspeed=50, lanes=2, sidewalk=both, height=24. OpenStreetMap calls these tags; commercial datasets call them attributes. They are what a machine actually reads.
• Topology. Lines that meet at a junction must share a node, so a router knows the roads connect. Lane-level maps also store which lane leads to which lane across a junction.
• Known accuracy. A sidewalk robot needs the kerb position to within decimetres. A fleet dispatcher is happy with the right street. The data must say which one it is.
• Freshness. Every object should carry a date, and the map needs a way to receive live changes on top of the stored base.

The automotive world has a standard picture for the last point. The ETSI technical report on the Local Dynamic Map (TR 102 863) sorts the data a connected vehicle uses into four types:

The useful idea is the layering. A base map changes slowly and can be shipped ahead of time. Everything above it arrives over the network while the vehicle drives. The same split works for robots, drones and fleets.

How much machine detail open map data has: our measurement

OpenStreetMap is the one large map whose full history is public, so it lets us measure how machine-relevant detail has grown. We used two sources:

• Europe, 2016 to 2026. The taginfo service for the Geofabrik Europe extract publishes a daily chronology of every key: how many objects gained or lost it. Adding up those daily changes gives the number of objects carrying each tag on any date. We took 1 January of each year and the latest day, 20 September 2026.
• Eight city centres, 2016 and 2026. With the ohsome API from HeiGIT, which queries the full OpenStreetMap history, we measured the share of car-road length that carries a speed limit, a lane count and a sidewalk tag, in a 4 km circle around the centre of each city, on 1 January 2016 and 1 July 2026.

Three findings stand out.

The pedestrian layer grew fastest. Kerbs, drawn sidewalks and sidewalk tags on roads grew between 9 and 83 times. Three newer keys that record the sidewalk on each side, sidewalk:left, sidewalk:right and sidewalk:both, went from under 1,300 objects combined in 2016 to 1.46 million. This is the layer delivery robots and accessibility apps need, and ten years ago it barely existed in open data.

Car attributes grew faster than the road network. Objects with a highway tag, which covers roads, paths and features such as crossings, grew 2.5 times. Speed limits grew 3.0 times and lane counts 3.7 times, so a larger share of the road network now carries them.

Height is still thin. About 10% of European buildings in OpenStreetMap have a level count. The height tag sits on 5.5 million objects, including towers and trees, which is at most 2% of the 253 million buildings. A drone planner or a 3D model still needs another source for most buildings, such as national lidar or 3D building programmes.

City centres: from half-mapped to nearly complete

In the median city centre, the share of road length with a speed limit went from about half to 95%. Sidewalk information went from 6% to 72%. Paris shows the size of the change: 5% of its central road length had a sidewalk tag in 2016, against 95% in 2026.

Lane counts stay lower everywhere, and that is partly by design. Many mappers leave lanes off residential streets, where software can assume two lanes, or one on a one-way street. A router that falls back on such defaults still works. A robotaxi cannot, and it uses its own lane-level map.

How maps for machines are made and kept fresh

No single method produces a map that machines can trust. Map makers combine several, each with a different cost and update speed:

• Surveys and mobile mapping. Vans with cameras and lidar drive the streets and capture lane markings, signs and kerbs with centimetre accuracy. It is the most precise method and the most expensive per kilometre, so it is repeated rarely.
• Aerial and satellite imagery. Good for geometry such as buildings, road outlines and sidewalks. Imagery shows a road but not its speed limit.
• Probe data from vehicles and phones. Millions of GPS traces show where traffic actually flows, how fast, and when a road closes. This is how live layers and travel times are built. Map matching is the step that snaps those traces to the road network.
• Government data. Road authorities hold the official speed limits, weight limits and roadworks. EU rules now require them to publish more of it, as the next section shows.
• Community edits. OpenStreetMap contributors add the local detail no fleet of vans reaches: the kerb ramp, the new bike lane, the shop that moved.
• Machine learning. Models detect signs in street-level images, buildings in aerial images and changes between two dates. They speed up every method above, and a human or a second source still checks what they find.

The layers age at different speeds. Building outlines last for decades, speed limits change every few years, roadworks change weekly and traffic changes by the minute. A map for machines keeps each layer on its own update cycle.

The European angle: privacy, data access and open data

Europe regulates most of the inputs above, and several rules push map data towards machines.

• Road data on National Access Points. Delegated Regulation (EU) 2022/670 on real-time traffic information has applied since 1 January 2025. Road authorities and operators must share the data they collect, such as the number of lanes, speed limits, access restrictions, height and weight limits and delivery areas, through a national access point in each Member State, in standard machine-readable formats such as DATEX II or TN-ITS. The regulation covers the whole public road network. Until the end of 2027, the duty for road network attributes such as lane counts covers only the trans-European network, other motorways and primary roads.
• Speed assistance in every new car. Under the General Safety Regulation (EU) 2019/2144, intelligent speed assistance has been required for new vehicle types since 6 July 2022 and for all newly registered vehicles since 7 July 2024. Article 6 lets the system take speed limits from road signs, from infrastructure signals or from electronic map data, or from both.
• Drone zones in a digital format. Implementing Regulation (EU) 2019/947, Article 15, requires Member States to publish their drone geographical zones "in a common unique digital format". The U-space rules in Regulation (EU) 2021/664, applicable since 26 January 2023, make a geo-awareness service with the latest airspace constraints mandatory in U-space airspace.
• Open geospatial data. The high-value datasets regulation (EU) 2023/138 requires public bodies to publish addresses, buildings, administrative units, cadastral parcels and geographical names free of charge and through APIs, since 9 June 2024.
• Access to connected-product data. The Data Act (EU) 2023/2854 has applied since 12 September 2025. Connected products placed on the market after 12 September 2026 must be designed so users can access the data they generate, and that includes vehicle location data. Our Data Act explainer covers what that means for location APIs.
• GDPR on probe data. A GPS trace from a car or phone usually relates to an identifiable person, so building a live map from probe data is processing of personal data. Map makers aggregate traces and drop identifiers before a trace becomes a traffic speed. Our guide to GDPR-compliant map APIs covers the processing side.

What this means for product teams

If a machine reads your map, design the product around attributes and the age of each value.

• List the attributes your machine needs, then measure coverage in your area. The city table above shows how much it varies: lane counts cover 73% of central Madrid's road length and 36% of Amsterdam's.
• Separate the static base from live layers. Ship the base map with the app or cache it, and stream closures, restrictions and traffic on top, as the Local Dynamic Map model does.
• Store the date of every value you rely on. When a speed limit or an opening time is older than your threshold, treat it as unknown.
• Use official sources where the law now provides them. National access points and high-value datasets are free, and they cover the attributes communities find hardest to collect.
• Give AI agents structured answers. A language model guesses addresses and distances. A map MCP server lets it call a geocoder or a routing API and get a number back instead.

Reproduce the measurement

The city measurement is one request per city and tag. This one returns the share of car-road length within 4 km of central Paris that carries a sidewalk tag, in 2016 and 2026:

curl -X POST "https://api.ohsome.org/v1/elements/length/ratio" \
  --data-urlencode "bcircles=2.3522,48.8566,4000" \
  --data-urlencode "time=2016-01-01,2026-07-01" \
  --data-urlencode "filter=type:way and highway in (motorway,trunk,primary,secondary,tertiary,unclassified,residential,living_street)" \
  --data-urlencode "filter2=type:way and highway in (motorway,trunk,primary,secondary,tertiary,unclassified,residential,living_street) and (sidewalk=* or sidewalk:both=* or sidewalk:left=* or sidewalk:right=*)"

The ratio field in each result is the share for that date. Our script also includes the motorway and road link classes, which makes no practical difference in city centres. For the European totals, call https://taginfo.geofabrik.de/europe/api/4/key/chronology?key=kerb and add up the daily changes.

When you build on this data, MapAtlas serves the same OpenStreetMap-based road network through its routing, matrix, isochrone and map matching APIs, processed in the EU.

Method and limits

• Europe counts: taginfo chronology of the Geofabrik Europe extract, data until 20 September 2026 for the chronology and 1 October 2026 for the current statistics. The two agree within 1% for every tag except sidewalk:both (4%), which grows fast in the 11 days between them. Counts are objects (nodes, ways and relations), so a long road split into several ways counts several times.
• City shares: ohsome API, elements/length/ratio, OpenStreetMap history until 27 July 2026, 4 km circles around each city centre, car roads from motorway to residential plus living streets and link roads. Dual carriageways count twice.
• What the numbers do not measure: whether a value is still correct, attributes that software infers from defaults (such as a national urban speed limit), the accuracy of geometry, and commercial map datasets, which are not public.
• Script: scripts/seo/blog-keyword-research/maps-for-machines.mjs in our repository; the full output is in the CSV below.

Sources

• ETSI TR 102 863 V1.1.1 (2011-06), Local Dynamic Map: rationale for and guidance on standardization
• Commission Delegated Regulation (EU) 2022/670, EU-wide real-time traffic information services, Articles 1, 3, 4, 14 and 16 and Annex
• Regulation (EU) 2019/2144, General Safety Regulation, Articles 6 and 19 and Annex II
• Implementing Regulation (EU) 2019/947, Article 15, and Implementing Regulation (EU) 2021/664 on U-space
• Implementing Regulation (EU) 2023/138, high-value datasets
• Regulation (EU) 2023/2854, Data Act, Article 50
• taginfo, Geofabrik Europe and the ohsome API by HeiGIT, © OpenStreetMap contributors