Self-driving progress needs better AI, not more sensors, says Tesla AI chief

Ashok Elluswamy, Vice President of AI at Tesla, said the main barrier to autonomous driving is artificial intelligence, not hardware sensors. Speaking at a conference, he argued that modern camera systems already capture enough data for cars to drive themselves. The difficulty, he said, lies in interpreting that data correctly.

“The self-driving problem is not a sensor problem, it's an AI problem,” Elluswamy said. “The cameras have enough information already. It's a problem of extracting the information,

which is an AI problem.”

He said Tesla’s system relies primarily on camera input and neural networks rather than multiple specialised sensors. According to him, the company uses an end-to-end AI model that converts raw video, vehicle motion data, and navigation inputs directly into driving actions such as steering and braking.

Elluswamy said this approach avoids the limits of traditional modular systems, where perception, prediction, and planning are handled separately. He said real-world driving situations are too interconnected for rule-based or segmented software designs.

He also showed examples where Tesla vehicles reacted to unusual road situations, including animals crossing and sudden vehicle crashes ahead. He said these responses require understanding intent and motion, which he described as AI inference rather than sensor detection.

He also said Tesla gathers rare driving scenarios from its global fleet to train its models. Most driving data is routine, he noted, so the company filters for uncommon or risky events to improve safety performance.

Elluswamy described Tesla’s use of neural networks that model the 3D environment around the car and simulate future scenarios. He said these systems help both training and safety evaluation by replaying past incidents and generating new test situations.

When asked why Tesla focuses on cameras instead of adding more sensors or roadside infrastructure, he said humans navigate using vision and that autonomous systems should solve the same core perception problem.

“It's so obvious you can solve this with cameras,” he said. “Why wouldn't you solve with cameras? It's 2026.” He added that earlier autonomous systems relied on extra sensors because AI was not advanced enough at the time.

“Nowadays, intelligence has grown tremendously. You can obviously extract this information, and once you have the AI, you don't need the other sensors.”