Tesla Builds Its Own Maps to Fix FSD Flaws

Tesla is developing its own internal mapping system to address persistent issues in its Full Self-Driving software. The company’s head of AI, Ashok Elluswamy, outlined this new direction during the Q3 2026 earnings call. The primary goal is to improve how the vehicle handles speed limits, parking maneuvers, and route selection. This move represents a significant pivot for a brand that previously dismissed the necessity of detailed maps for autonomous driving.

The decision follows a direct question from a retail investor during the shareholder Q&A session. The investor noted that mapping remains a core weakness of the system, specifically citing problems with speed limit recognition, parking, and typical route choices. They asked why Tesla continued to rely on third-party data providers. Elluswamy agreed with the premise, stating that the company should not depend on external entities for this critical data. He confirmed that Tesla is now constructing a proprietary map using a new AI-based architecture.

This new system will be fed by the collective learning of the Tesla fleet. As vehicles drive, they gather data that helps build a shared map. The system is also designed to learn individual driver preferences and respond to voice commands. Elluswamy indicated that the technology will initially focus on the beginning and end of trips with FSD version 15. Over the coming months, it is expected to take over broader navigation, parking, and destination handling tasks.

This shift marks a return to a concept Tesla explored earlier in its history. When Autopilot launched in 2015, the company told Fortune it was building high-precision maps with significantly more granularity than standard navigation tools. However, at its 2019 Autonomy Day, CEO Elon Musk reversed this stance. He described high-precision lane maps as a mistake and a really bad idea that made systems brittle. In that same presentation, he famously called lidar a fool’s errand, arguing that Tesla could solve driving with cameras and AI alone.

Seven years later, the strategy has evolved. Tesla is now using its fleet to build a map layer to correct errors that its vision system frequently makes. To be clear, this is not the same type of high-definition map that Musk criticized in 2019. Tesla has always used standard navigation maps, so creating its own is not a contradiction in itself. However, the specific problems being addressed, such as speed limits and parking, require a deeper understanding of the environment than basic navigation provides.

The urgency of this fix is highlighted by recent performance data. Last month, a Belgian road safety group found that FSD vehicles sped in 55% of 30 km/h zones in Brussels. The average speed recorded was 44 km/h, partly because the car’s display showed an incorrect 50 km/h limit. In the United States, the NHTSA is weighing an investigation into claims that Tesla Robotaxis speed through active school zones. This issue covers about 1,000 Model Y and Cybercab vehicles.

If an end-to-end vision system struggles to read a simple speed limit sign, and the solution involves better map data, the vision-only approach may not be sufficient. This brings Tesla’s methodology closer to that of Waymo. Waymo maps cities by driving sensor-equipped vehicles through them before launching service and keeps those maps updated. Tesla is now describing a similar process where fleet data feeds a central map. The main hardware difference remains that Waymo vehicles are equipped with lidar.

This approach also aligns with how Tesla’s Robotaxi service operates. The service runs city by city within small geofenced areas. While Tesla is not currently advertising centimeter-level high-definition maps, a fleet-built map that becomes more detailed over time could easily move in that direction. There is nothing inherently wrong with using maps, as every serious self-driving program does. A map built from millions of Tesla cars could be a competitive advantage.

The challenge lies in the messaging history. Musk sold FSD as a general solution that worked anywhere through vision alone, while dismissing other methods as crutches. Now, with the addition of a map layer and geofenced operations, the distinction has blurred. Unsunsupervised FSD for consumer cars is currently targeted for Q2 2027. As the decade of self-driving efforts continues, both companies appear to be solving the same problem in very similar ways, with the only major difference being the presence of lidar on one fleet.