Tesla took an early lead in the race to develop self-driving cars, and its Autopilot system remains state-of-the-art. But the technology has advanced more slowly than the company expected, and Elon Musk's promise of a driverless demo drive across the US in 2018 has yet to materialize. Meanwhile, competitors are hard at work developing their own self-driving technology. GM's Super Cruise is currently available on the CT6 luxury sedan.
Is Tesla in danger of falling behind in the self-driving race? Medium contributor Trent Eady takes a detailed look at the company's Autopilot technology and argues that the Californian automaker will continue to lead the way.
All Tesla cars built since October 2016 are equipped with a suite of hardware designed for full self-driving, including cameras, radar, ultrasonic sensors and an upgradeable on-board computer. There are currently around 150,000 of these “Hardware 2” Teslas on the road, and they could theoretically be upgraded to self-driving cars via an over-the-air software update.
Above: As it stands, Tesla's Autopilot requires a hands-on approach (Youtube: Tesla)
Tesla disagrees with most others in the self-driving industry when it comes to LiDAR, a technology that uses pulses of infrared laser light to calculate distance. Companies like Waymo, Uber and others appear to view LiDAR as an essential component of their self-driving systems. But Tesla's Hardware 2 sensor suite doesn't include LiDAR, relying instead on radar and optical cameras.
Lidar's strength is its spatial accuracy: it can measure distances much more accurately than current camera technology. (Eady believes that better software could make cameras more accurate.) Lidar's weakness is that it doesn't work well in bad weather: heavy rain, snow, and fog refract and scatter lidar's laser pulses. Radar works much better in harsh weather conditions.
According to Eadie, cost may be the reason Tesla is avoiding lidar: “Self-driving grade lidar is prohibitively expensive, so Tesla could never put it in a production car. As far as I know, no affordable self-driving grade lidar products have been announced yet, and that seems years away.”
If Elon Musk and his self-driving team are convinced that lidar is unnecessary, why is everyone else so convinced that it is? “Lidar has an aura of magic in the public's imagination,” says Eadie. “It's easy to accept the new and incredible idea of a self-driving car if you explain that it's enabled by cool, futuristic laser technology. On the other hand, it's hard to accept the idea that you can build a car that can navigate complex city streets by connecting a plain old camera to a deep neural network.”
These deep neural networks are the real reason Eady thinks Tesla will stay ahead of its competitors in self-driving: The massive amounts of data Tesla is collecting through sensors in its existing fleet of roughly 150,000 Hardware 2 vehicles “provide real-world testing and training on a scale never before seen in the history of computer science.”
Rival Waymo has a computer simulation with 25,000 virtual cars, generating data from 8 million miles of simulated driving per day. Of course, Tesla's real-world data is far more valuable than any simulated data, and the company feeds it into deep neural networks to continually improve Autopilot's capabilities.
Deep neural networks are a type of computing system loosely based on the structure of the human brain (sounds like the kind of AI Elon Musk is worried about, but we have to trust that Tesla has it in control). Deep neural networks are good at modeling complex nonlinear relationships. The more data available to train the network, the better it will perform.
“Deep neural networks started to gain popularity in 2012 when a deep neural network won the ImageNet Challenge, a computer vision competition focused on image classification,” Eady explains. “In 2015, a deep neural network first narrowly beat the human benchmark in the ImageNet Challenge. The fact that computers can even outperform humans at some vision tasks is exciting for anyone who wants to get computers to do things better than humans – like driving.”
By the way, who was the human benchmark beaten by a machine in the ImageNet Challenge? Andrej Karpathy, who is now Director of AI at Tesla.
===
Note: This article was originally published by Charles Morris on evannex.com. Source: Medium
