Tesla Says It Will Build Its Own Maps for FSD, Seven Years After Musk Called High-Precision Maps 'a Huge Mistake'

Tesla's camera-only driving system is going to carry maps after all, and the person who said so is Ashok Elluswamy, the company's vice president of A.I. software. Asked on the third-quarter earnings call why Tesla relies on third-party maps for speed limits, parking and typical routes, he answered that "we shouldn't be depending on external people's maps", according to Electrek, which reported the exchange on October 22.

What Elluswamy described

The question came from a retail investor in the shareholder Q&A, who said mapping remains a problem for Full Self-Driving. Elluswamy said Tesla is building its own maps through a new A.I.-based mapping and navigation architecture, fed by fleet learning. The system will also learn individual drivers' preferences and respond to voice commands.

The rollout is sequenced. Electrek reports that FSD v15 will begin handling the start and end of trips, with navigation, parking and destination handling taken over in the months after. No map coverage figure, fleet size or ship date for the full stack was given.

The record Tesla is reversing

Electrek lays out the history. In 2015, when Autopilot launched, Tesla told Fortune it was building high-precision maps with "100 times the level of granularity" of standard navigation, using customer-car data. At Autonomy Day in 2019, Elon Musk called high-precision lane maps "a huge mistake", relying on them a "really bad idea" that makes systems "extremely brittle", and described lidar as a "fool's errand". Those quotes are Electrek's account of the event; this article did not review the 2019 recording.

Date	Tesla position on maps	Source
2015	Building maps 100 times as granular as standard navigation	Fortune, as cited by Electrek
2019	High-precision maps are "a huge mistake"	Musk at Autonomy Day, as cited by Electrek
October 21, 2026	"We shouldn't be depending on external people's maps"	Elluswamy on the Q3 call

The reversal has a narrower reading than Electrek's. Musk's 2019 objection, as Electrek reports it, was to high-precision lane maps. Elluswamy described fleet-learned maps with an A.I. layer, which is a different construction. Electrek argues the difference does not rescue the earlier messaging that vision alone would work anywhere, and notes Waymo also maps cities, though with lidar-equipped vehicles.

Why the sensors matter to the map

A map built from fleet data inherits the fleet's sensors. Tesla's messaging has been that vision alone would work anywhere, so the map can hold what a camera sees: signs, lane geometry, driveways, parking lots. It cannot hold depth measured by a laser, which Waymo's map-building cars carry. Whether that is enough is the empirical question, and the public tests so far are not reassuring on one point the maps are meant to fix.

A Belgian road-safety group tested FSD in Brussels, and, per Electrek, the car sped in 55% of the 30 km/h zones it drove, averaging 44 km/h, while its screen often showed a 50 km/h limit. Speed limits are one of the problems the investor on the call named, so the Brussels result is the nearest public measurement of what the new maps are meant to fix.

What the sequencing implies

The first job Tesla gives its own maps is the two ends of a trip. That is where the investor's examples, parking and typical routes, live, Handing those to FSD v15 first, then navigation and destination handling in the months after, means the map layer is being built from the edges of the journey inward.

Electrek's point is that the fleet is already doing this work: every Tesla driving a route is surveying it, much as Waymo's vehicles survey the cities where they run. The difference Electrek draws is the sensor, not the principle. Electrek's account does not name the third-party maps Tesla uses today or say how much of its navigation would be replaced, so the size of the change is unknown.

The schedule: unsupervised FSD at Q2 2027

The TechCrunch account of the earnings call puts the maps answer in context. Elluswamy said the first customer vehicles could join the robotaxi fleet around Q2 2027, when owners' cars would get a software update enabling unsupervised driving. In April, Musk said the unsupervised version would "probably" arrive in Q4 2026. Today owners can subscribe only to the supervised version, which requires the driver to stay alert.

The regulatory side is open. Electrek reports that NHTSA is weighing an investigation into claims that Tesla's robotaxis speed through active school zones, with the covered fleet at about 1,000 Model Y and Cybercab vehicles. The robotaxi service itself operates in small geofenced areas.

What would change my read

The claim to test is that fleet-learned maps close the gap between supervised and unsupervised driving. Three disclosures would settle it: the share of trips covered by Tesla's own maps, a speed-limit error rate measured the way the Brussels group measured it, and whether the Q2 2027 date survives the first v15 release. The denominator is the one that matters; a 1,000-vehicle robotaxi fleet is a small sample against the cars that would receive the update.

Related hardware coverage: Porsche's wireless charging at 11 kW and Rio Tinto's 360 MWh heat battery.