Could a line of unmanned Teslas roaming Austin streets represent a turning point for autonomous mobility, or a regulatory showdown? This past weekend, Tesla CEO Elon Musk confirmed that “testing is underway with no one in the car” to mark the brand’s first unmanned autonomous vehicle trials in Austin, following six months of manned robotaxi pilots. This latest news sent Tesla stock up by 3.5% to a close of $475.11 on Monday, a market close for 2025, representing the highest close for Tesla all year long thus far.
Starting in June, the Austin pilot started with safety drivers in the front passenger seat, followed by handing them the wheel. Now, with human safety monitors out of the loop, Tesla is fully depending on its Vision-based AI software stack, known as Tesla Vision, using eight cameras and its proprietary HydraNet neural network design capable of simultaneously performing 50 compute tasks on its Full Self-Driving (FSD) computer. While most otherAutonomous Driving System (ADS) startups and tech firms are incorporating multiple sensor technologies like lidar and radar in their hardware software framework, Tesla is taking a unique approach, using its camera inputs to generate Bird’s Eye View spatial maps, allowing estimates of depth, lane detection, and road graph prediction. Training these multiple-head networks is quite compute-heavy; Tesla uses distributed PyTorch pipelines and its in-house Dojo supercomputer.
Safety, however, is still an area of concern. There were seven collisions reported by Tesla mid-way through October among its Austin vehicles, all featuring the new automation systems. Though non-serious, Carnegie Mellon’s Philip Koopman observed: “with such a small fleet, there should have been fewer than seven reportable accidents, especially considering that there is a safety supervisor in each one whose job is to prevent crashes.” The situation is similar for AVs. Analyses by the California DMV have shown that the majority of AV accidents are rear-end impacts, accounting for up to 69 percent of reported collisions, mostly occurring when AVs are stationary at intersections.
Tesla chose to excise “narrative descriptions” from its reporting to the National Highway Traffic Safety Administration, and because of this, there is less transparency about the context of incidents, and policymakers are no longer tolerating this circumstance. In Texas, it is legal without permits, but Senate Bill 2807 will, starting May 2026, make it necessary to have authorization from the DMV and a safety plan for first responders when deployed commercially. In the state of California, a more intensive regulation requires multiple permits from both the DMV and Public Utilities Commission, of which Tesla does not yet have.
The competitive environment is becoming even more challenging. Alphabet’s Waymo is already hosting fully autonomous vehicles in five U.S. cities and has recorded more than 14 million miles in 2025 with the data indicating an 79% reduction in airbag deployment incidents relative to human drivers. WeRide and its competing service for the region, Apollo Go, are growing very quickly with the latter company already providing 3.1 million fully autonomous rides in the third quarter of 2025 with plans for expansion in Europe and the Middle East. Both competitors use sensor-rich platforms like the RT6 from Apollo Go, with almost 40 sensors on board including eight lidar. This is in contrast to the hardware minimalist nature of Tesla.
From an perspective, Tesla’s Austin testing will put their vision-only ADS to the test under uncontrolled urban conditions without the safety net provided by the human presence onboard. The ability to cope with “edge cases,” which are rare and difficult to forecast instances of human behavior (such as police hand signals and erratic pedestrian movements), demands excellent perception and planning redundancy. Highly advanced simulation infrastructures like scenario-based simulation and sensor-in-the-loop simulation have been industry norms for ADAS validation before the commencement of on-road testing and enable the accumulation of simulated miles overnight on a scale of millions. Whether Tesla’s validation loop encompasses the same scope of simulation data as their competitors will have to wait to be answered.
Musk plans to double the number of vehicles in Austin to 60 by end of year, which was nowhere close to the estimate of 500. The deployment rate indicated can be seen in Tesla’s own post when they said “Slowly then all at once.” The oncoming achievement for those familiar with technology and involved with EVs will mark a notable point of operating driverless vehicles; however, its feasibility will depend on safety and Tesladeployment.
