Tesla: Self-driving simulation, lidar give 'false sense of progress'

Elon Musk predicts other companies will dump lidar once they "solve the vision problem."

Tesla has shed more light on Autopilot's neural-network prowess, arguing that its superior machine-learning methods eliminate the benefits of simulated training and lidar sensors.

The company's Autopilot software head, Andrej Karpathy, has explained how the entire fleet of Autopilot-equipped cars contributes to an automated training system that has allowed the neural-network algorithms to become more refined with little need for computer simulation.

Karpathy and Tesla CEO Elon Musk argued that simulation does not capture the "long-tail" rare events that prove most troublesome for self-driving platforms, such as complex pedestrian zones where the pedestrians may be moving in reaction to the car's motion.

"If someone can make a simulation that accurately matches reality, that in itself would be a monumental achievement," Musk said. "They can't; there's no way."

The company also uses data from cars that have Autopilot turned off to help train its neural network, using the human driver's behavior as an example of how to maneuver in certain situations. A demonstration video showed how this information is used to improve predictions for lane guidance that follow a curve beyond the line of sight.

Tesla has reaffirmed its commitment to using cameras and radar/ultrasonic sensors for depth perception, pointing out that vision is proven to be an effective sense used by humans to drive. Autopilot uses stereo imaging from multiple cameras, along with raw neural-network learning to process motion into depth. In a short six-second clip of a drive in a city, the system created a 3D point cloud of the surrounding vehicles, pedestrians and buildings.

Musk argues that lidar gives a "false sense of progress" for companies that have not solved "the vision problem" through neural-network programming alone.

"Lidar is a fools errand," he proclaimed. "Anyone relying on lidar is doomed."

Nearly 500,000 Autopilot-equipped cars are already on the road, with another 500,000 expected to be added in the next year. Tesla believes its unmatched real-world data set gives the company an edge above all rivals.

Waymo is widely considered the leader in full self-driving development, though the company has cautiously launched its commercial service only in a small patch of the Phoenix suburbs. Tesla's system still requires drivers to keep their hands on the wheel, but the company predicts this precaution will be overcome in the next year as the ongoing machine-learning programs continue to refine the software.