Tesla: Vision Only

By mimicking humans, who navigate mostly through vision, Tesla argues that a high-resolution camera system – backed by sufficient computing power, AI, and data – is superior to a suite of sensors that may provide conflicting data (sensor noise).

Tesla has doubled down on vision-only, removing radar and ultrasonic sensors in favor of an “end-to-end” neural network. Eight cameras, providing 360-degree visibility, capture raw visual data processed in real-time by an onboard computer powered by Tesla-designed AI silicon. Instead of using radar to measure distance or velocity, Tesla’s deep learning models infer the depth, speed, and geometry of objects purely from 2D video feeds. By training on billions of miles of real-world driving data, the system learns to navigate complex maneuvers and rare “edge” or “long-tail” scenarios that are virtually impossible to program with traditional hand-coded rules.

Tesla Isn’t Waiting

Tesla isn’t going to wait for perfection (it will never happen anyway). It’s pushing ahead – too quickly, critics claim – with vision-only Full Self-Driving (FSD). But there is method to its madness. While driverless (Level 4) Waymo service is already available in over 10 U.S. cities in geofenced areas, the cost of a Waymo taxi – retrofitted with LiDAR, radar, cameras, and microphones – is estimated to be tens of thousands of dollars higher than a Tesla. So, basically, Tesla is trying to achieve the same driverless goal as Waymo but at a cost the consumer can afford. “Tesla developed high resolution radar and the hardware is actually present in Model S & X, but it just can’t compare to passive optical (cameras), so we turned it off,” CEO Elon Musk said in August 2025. “We certainly hope to have unsupervised FSD/Robotaxi operating in, I don’t know, a dozen or so states by the end of this year,” Musk said in the first quarter earnings conference call this past week (April 22).

Which Approach Wins?

“Technically, using multimodal data rather than camera-only data makes complete sense,” Jason Corso, chief scientist and cofounder at Voxel51 and Toyota Professor of AI at the University of Michigan, told me in an email, arguing in favor of Rivian’s approach. “One might argue that humans can drive fine with only their eyes, but the adaptability of the human visual system puts camera systems to shame,” he said, adding, “Frankly I think the cost argument is baseless. At scale, the cost of cameras and LIDAR would drop commensurately.”

Philip Koopman Emeritus Professor, Electrical and Computer Engineering, at Carnegie Mellon University, pointed me to an October 2024 post (he didn’t offer a comment beyond that). “Tesla has a huge fleet size. They can’t afford to put expensive sensors in each vehicle due to cost pressure if they want to make a profit,” he wrote. “Trying to scale other techniques…is going to run into problems with edge cases…The path to avoiding those problems guides Tesla to an end-to-end (E2E) reinforcement learning approach,” he wrote.

Koopman goes on to say, however, that robotaxi success depends on “acceptably safe” performance during high-consequence, rare “edge cases.” While Tesla’s massive fleet provides a data volume advantage for training, he notes their advantage is diminished by the use of “mediocre cameras” rather than the higher-quality sensors used by competitors, making Tesla’s billions of miles of data less effective for machine learning. (See latest Tesla cumulative FSD miles here, which is trending toward 10 billion miles. It should be noted that newer Model 3 and Model Y vehicles reportedly feature higher-resolution cameras.)

Rivian And Tesla Have Same Goal, Different Timeline

Rivian has the same goal as Tesla but with a different timeline. “Our view and really strong conviction is that over the course of the remainder of this decade…as we start to move to hands-off and eyes off and then ultimately to Level 4 where the vehicle can operate itself entirely on its own,” said Rivian CEO RJ Scaringe in the February earnings conference call.