What HD Maps Actually Contain, and Why They Go Stale

A turn-by-turn navigation app only needs to know which street you're on. An automated-driving system needs to know which lane, where the curb starts, and where the paint on the road actually is, down to the centimetre.

That gap in precision is the whole reason HD maps exist as a separate category from the maps most people use every day. An HD map used in automated driving typically encodes lane-level geometry, curb positions, sign and signal locations, and the exact boundaries of the drivable surface. Consumer navigation data gets away with far less: road-level accuracy is enough to tell a driver where to turn, and lane-level detail would be wasted on a human reading a screen anyway. A vehicle deciding whether it can safely edge toward the shoulder to let a stalled car pass, on the other hand, needs to know precisely where the drivable surface actually ends, and that's information a standard consumer map was never built to carry.

The density is also the source of the maintenance problem. Roads change constantly. Construction crews shift lane markings for months at a time, cities repaint intersections, new signage goes up and old signage comes down, and every one of those changes can make a previously accurate HD map wrong in a way that matters to a vehicle relying on it. Re-surveying isn't a one-time job. It has to run continuously, and an out-of-date HD map is a documented source of automated-driving error rather than a theoretical risk. A map that hasn't caught up with a repaving project can still show a lane in a place it no longer exists, which is a far bigger problem for a system trusting centimetre-level geometry than it would ever be for someone glancing at a phone screen.

The map's second job, beyond simply describing the road, is localization. A vehicle doesn't just consult an HD map to see what's ahead; it continuously compares what its own sensors are detecting against the map's stored geometry to figure out exactly where it sits within a lane, often to a precision no GPS signal alone could provide. GPS alone can drift by several meters in ordinary conditions, which is fine for telling a human which street to turn onto and nowhere near good enough for deciding whether a vehicle's wheels are centered in a narrow lane. The HD map effectively becomes a second reference frame the vehicle checks itself against many times a second, which is also why a map that's subtly wrong doesn't just mislead the route; it can throw off the vehicle's own sense of where it is.

Keeping that reference current is a coverage-versus-frequency tradeoff as much as a technical one. A dedicated survey vehicle can capture extremely precise geometry but can only cover so many roads before the data it collected starts to age. Crowdsourced updates, drawing on discrepancies flagged by a much larger fleet of ordinary vehicles already driving those roads, can catch changes faster and across a wider area, but at lower per-observation precision than a purpose-built survey pass. Most mapping providers end up running some blend of the two rather than picking one method outright, precisely because neither approach alone covers both the precision and the freshness an automated-driving system actually needs.

TomTom, one of the D20 index's mapping constituents, is built around exactly this re-surveying problem at scale, and its specific approach to keeping automated-driving maps current is covered separately in the breakdown of TomTom's HD mapping methodology. A single repainted intersection or a temporary construction lane shift is enough to invalidate geometry a planning system was relying on, until the next survey pass catches up.