Tesla’s Autopilot Is Misleading, Say Tesla Engineers

Elon Musk stubbornly clings to cameras, ignoring radar and LiDAR

A Tesla using Autopilot on a Los Angeles freeway. (Picture courtesy of New York Times.)

A Tesla using Autopilot on a Los Angeles freeway. (Picture courtesy of New York Times.)

UPDATE Jan 18, 2022 – The state of California is
charging the driver of a Tesla with two counts of vehicular manslaughter after a
2019 accident when his vehicle using Autopilot ran a red light coming off
a highway near Los Angeles, according to an

NPR news report
. It is second time a state has leveled charges
for the use of self-driving technology. Arizona charged an Uber “driver” for her
role in the death of a woman crossing the road after one of its experimental
self-driving vehicles

to recognize the danger in 2020. The California case is
the first involving self-driving technology in a production vehicle.

In his race to be the first to have a true self-driving car, Elon Musk is ignoring his company’s own engineers, according to a recent report by the New York Times.

Tesla Engineers have implored Musk to use more than just vision systems (cameras) as sensors on the vehicles, but according to the engineers, Tesla CEO Elon Musk has vehemently opposed the addition of technologies widely believed to enhance the vehicles’ detection of objects, such as radar, ultrasonic sensors, radio waves and LiDAR.

“We can see with just our eyes,” Musk has been reported to say, going back to the very first principles of our biology. It does not seem to bother him that our eyes can be blinded by glare, can’t see through smoke and fog—both all too common in California where Tesla vehicles are produced.

Musk has been stubbornly clinging to the belief that a vision system similar to human stereoscopic vision is all that a self-driving vehicle needs and has ordered his engineers to produce such a vehicle. But the vision we take for granted as adults took years to develop over our infancies. The signal processing our brain performs to interpret electrical signals from our eyes as we react to objects is no trivial system to duplicate. Having to interpret the position of objects from grainy, black and white images with stereoscopic vision introduces an extra step and is not as accurate as LiDAR, which returns exact XYZ coordinates.

Other companies, like Mercedes, arguably the most safety-conscious of all the big car companies, and Tesla’s most recent competition, Lucid, use multiple sensor technologies. The Lucid Air Dream Edition (priced at $169,000) and the Lucid Air Grand (priced at $139,000) beat the Tesla Model S in performance (0-60 mph in a little over 2 seconds compared to 2.5 seconds for the Tesla Model S Plaid). Their range (520 miles vs 396 for the Tesla) may also trump Tesla in safety, sporting 5 radar sensors in addition to 14 visible light cameras, 4 “surround-view cameras” 12 ultrasonic sensors and, perhaps the most significant, a solid-state LiDAR unit.

The Lucid is the first consumer production vehicle sold in North America to include LiDAR. The ultra-luxury Mercedes-Maybach S-Class includes LiDAR but is sold only in Europe. LiDAR, for its expense and size, had been seen in the U.S. only on top of experimental fully autonomous vehicles as part of their “sensor fusion”—a pile of technology bolted onto their roofs.

Tesla is under investigation by the National Highway Traffic Safety Administration after a series of incidents when its vehicles using Autopilot drove into emergency vehicles, apparently confused by their flashing lights at night.

Musk has been quoted as saying his cars have all the hardware they need for Level 5 autonomy. A time out to consider Level autonomy, the zenith of self-driving, where there doesn’t need to be any hands on the steering wheel, a steering wheel, or even hands. With Level 5 autonomy, the car knows to get from point A to point B, through neighborhoods and over highways, avoiding collision with whatever may be on the road, after the passenger merely tells it where to go.

Saying that Tesla vehicles are hardware ready has caused confusion among Tesla owners, who jump to the conclusion that hardware is all that is needed. The U.S. government has sided with the owners who have reached this conclusion and has asked that Tesla not to use the language “fully self-driving” to describe its vehicles.

If the vehicle owners had continued to listen, they would have heard that it would take future updates to the vehicles’ software for them be counted as true Level 5 autonomous vehicles, as in this 2016 video statement.

Were Tesla only to be consistent. More recently, Musk has stated that Tesla’s Full Self-Driving (FSD) package—currently being beta tested by volunteer Tesla owners—will allow cars to drive themselves on city streets as well as highways, according to the New York Times article.

And that was when the U.S. National Transportation Safety Board (NTSB) became concerned with “the language that’s used to describe the capabilities of the vehicle,” said Jennifer Homendy, chairwoman of the NTSB, which has investigated accidents involving Autopilot. “It can be very dangerous,” she added.

If there is any hope of Musk changing his mind about adopting more technologies, it was suggested in an August 2021 tweet where he admitted that the “FSD Beta 9.2 is not great.”

What’s Wrong with LiDAR and Radar?

An experimental self-driving Mercedes S-class spotted outside the AWE conference in San Jose. Note the LiDAR unit on the vehicle’s roof.

An experimental self-driving Mercedes S-class spotted outside the AWE conference in San Jose. Note the LiDAR unit on the vehicle’s roof.

Short answer: LiDAR is too expensive and radar is fuzzy. However, the advantages of these technologies to autonomous vehicles are too great to ignore.

A LiDAR for automotive use is both ungainly (about the size of a bucket) and expensive. Inside a tilting and rotating mirror directs a laser from which a 3D point cloud is generated. The points are highly accurate in X, Y and Z coordinates. A laser/radar system can penetrate fog, smoke and the dark of night—unlike visible light cameras.

The LiDAR unit in use by Google and Uber’s experimental self-driving cars may cost as much as $80,000. Auto manufacturers may be fond of saying that the hardware needed for autonomous vehicles (LiDAR, radar, cameras, etc.) adds only $8,000 to $10,000 to the cost of a production vehicle, but the cost of the full setup bolted onto an experimental self-driving car was estimated to cost a whopping $250,000 in a 2017 Quartz article. Even at the lower end of the range, it will be hard to justify a safety option that costly for all but the most expensive vehicles, such as Teslas, Lucids, Mercedes and BMWs.