Geohot announced the "hacker" autopilot for $ 999

Prototype autopilot sample December 2015. The PC is installed in the Acura ILX glove compartment, in the cabin there is a 21.5-inch monitor in portrait mode (the Tesla display is only 17 "), a joystick near the parking brake to turn on the autopilot. Photo: Geohot

American schoolboy George Hots (Geohot) became famous throughout the world at the age of 17, when he hacked the iPhone in 2007. Three years later, he became the author of the Sony PlayStation 3 jailbreak. Then there was a showdown with Sony and work in several large companies that invited young talent: Google, SpaceX, Facebook. Hacker did not see anything interesting in these firms. He had a little fun at hacker contests for hacking browsers, earned several hundred thousand dollars in prize money, but this also quickly bored. Last year, Geohot still found a case for everyone. He founded comma.ai (“comma.ii”) and launched his own project: a homemade unmanned vehicle . For a couple of months, Geohot created an autopilot system for his Acura ILX, which is not inferior to the Tesla autopilot, and in some places even surpasses it. The system is fundamentally different from all existing on the market: it is a self-learning neural network that "observes" the actions of the driver.

Prototype

To connect to Acura ILX automotive systems, the hacker had to recall old social engineering skills. He filed an online application and registered as an authorized Honda service center. The application was approved - and the hacker gained access to reference guides and charts for Acura.


Intel NUC mini-computer, a pair of GPS modules, a router and other details of the Acura ILX car computer with autopilot. Photo: Geohot
')
The tracking system for the "hacker" autopilot consists of six cheap video cameras costing about $ 13 - about the same set in smartphones - and the lidar on the roof.



In a later version of the autopilot, a radar of a different model was used instead of a lidar, tied with tape to the grille .

The neural network in the first version of the autopilot was a program of just 2,000 lines of code. During the two months of training in a hacker’s car, the program learned how to drive perfectly on the highway, change lanes to lanes, brake and accelerate as needed, fit perfectly into turns, etc. In fact, this program predicts the actions of the driver. For example, if the autopilot sees a cyclist, then he frees him a little band, because the driver did so in the past.

Last year, Geohot rejected Ilona Mask's job offer at Tesla with a multi-million dollar bonus. Moreover, Geohot invited the lead engineer Tesla Motors to his startup Comma.ai. It seems that between geeks scheduled mezhdusoboychik.


The colleague's photo is set at the place of honor in the Geohot office. Photo: Chad McClymonds

Comma one

It has been almost a year since the first demonstration of the prototype self-made autopilot. Now this system is almost ready for market launch.

In August, Geohot posted a scientific paper on Github describing the autopilot neural network, source code and training data set (80 GB).

According to the author, 7.5 hours of training is enough for a neural network to master basic driving skills. After that, everyone will be able to repeat the experiment that George Hots showed last year to a Bloomberg journalist on his Honda.

Operation of individual autopilot software modules
Predictor Steering Wheel Training


Training model generating images


“Zoox, Tesla, Google and Cruise, we are waiting for your releases of your open source software,” said George to competing companies that are developing in this area. “Open software is better than closed.” By the way, the mentioned autopilot developer Cruise Automation was recently bought by GM for about a billion dollars .

Geohot says that now their program is not inferior in Tesla's capabilities. "If Tesla is like iOS for robotic cars, then we want to be Android."

Hacker invited everyone to participate in a crowdsourcing project. The Comma.ai website features a beta version of the Chffr mobile app, abbreviated from driver ( android , ios ), which turns a smartphone into a DVR. Attach a smartphone on the windshield - the program monitors your driving, records video with a small frame rate and sends the collected data via WiFi to Comma.ai servers. Perhaps this information will be used to train the neural network. Beta participants can view their travel statistics and video footage for each route on the Comma.ai website.


For each minute of use of the program, comma points are awarded, which will definitely be needed in the future. “These scores are absolutely unbelievable and you really want to have them,” said George Hots. “In a few months you will be happy that you have them.”


“In a few months you will be happy that you have comma points,” said Geohot

Yesterday at the TechCrunch Disrupt SF conference, the hacker held an official demonstration of the autopilot , which will soon go on sale under the name Comma One .


Geohot at TechCrunch Disrupt SF, September 14, 2016

Comma One is a plug-in hardware and software system for autonomous driving. Unlike Google and Tesla, you don’t need to buy a new car together. Equipment sold separately. For $ 999, a customer receives a 24-month software subscription.

Geohot said that installing an autopilot in his car would be as easy as assembling Ikea furniture. To install the autopilot requires that the car had at least an electric power steering and electronic stability control (ESC) system. Adaptive cruise control or automatic re-entry into the adjacent lane is not required.


Autopilot Show in May 2016

Let's hope that the commercial version does not need to wind the radar with adhesive tape to the grille. A few months left before the sale of Comma One. So far, we accumulate comm points. At first, Comma One will be available only for a small number of users and a limited list of car brands.

George Hots intends for several years to get ahead of Google, Tesla and other competitors in the development of automatic piloting systems for cars. By the end of 2018, he says, users should hit approximately 1.5 billion kilometers in the Chffr program. Training a neural network on the habits of many drivers will allow to correct the characteristic shortcomings of driving each of them.

“All good drivers are alike, and every bad driver is bad in its own way,” says George Hots. “A generic driver will be excellent.”


George Hotz with his engineers from Comma.ai