GeekPwn 2016 | Geohot: My driver assistance system is better than Tesla's

【Geohot】

Geohot, the "magic boy" who cracked the world's first iPhone at the age of 17, came to the GeekPwn (awesome) hacking competition held in Shanghai.

This time, in addition to showing his many fans his admiration, he also brought his most valuable treasure: the car assisted driving system.

Not long ago, Geohot, as the CEO of "Comma.ai", launched an autonomous driving product - Comma One in San Francisco.

According to Geohot, this assisted driving system has the following features:

1. The price is $999, and the monthly service fee is $24, which is a bargain in the assisted driving system. (At least you don’t have to buy a Tesla to experience assisted driving)

2. This assisted driving system can be applied to all models.

3. The functionality of Comma One is comparable to Tesla's AutoPilot, and will be even better in the future.

This amazing guy values ​​his painstaking work very much, and chose the hacker conference that matches his "background" - GeekPwn, the most popular hacker show in China - as his first roadshow after the release.

In an exclusive interview, Geohot told Leifeng.com about the basic working methods of Comma One:

At present, Comma One works very similarly to Tesla's AutoPilot, and can achieve automatic following, automatic turning, etc. However, in the future, its performance will definitely be better than Tesla's.

We used more than 500,000 human driving data to train the system, which Tesla's assisted driving system does not have. We also made many optimizations, such as:

• While driving, the driver can immediately disengage autonomous driving by stepping on the brakes or pressing the cancel button.

• When driving at high speed, the steering wheel will never make stupid moves such as sudden turns.

As for why the wonder boy didn't just continue his iPhone hacking career, but instead tried to steal the job of assisted driving? Geohot's answer was very frank:

Because Apple's security has improved a lot, and it doesn't need my help. But now it seems that car manufacturers are stupid. So far, they have made beautiful cars, but they are not smart at all. They need my help.

Geohot said that researching assisted driving systems is actually quite difficult.

There are rules on the road, but many Americans, and of course many Chinese, do not drive according to the rules. So in this case, we cannot drive purely according to the rules, but let the machine learn: what choices people will make in this situation.

My ideal situation is: we use 500,000 samples to do some AI learning in the early stage. After people use it, they will provide more data for us to improve the system, which will make the system smarter and more people will buy it. This is a virtuous cycle.

It seems that third-party assisted driving systems, represented by Geohot's Comma One, are ready to compete with the intelligent driving systems originally installed by automobile manufacturers.

However, with the promotion of third-party intelligent driving systems in the future, an foreseeable problem will arise, that is, safety.

For example, how reliable is the third-party driver assistance system? If a car has an accident, should the responsibility be borne by the car manufacturer or Comma One?

Geohot is not worried about this issue: I am very confident in the reliability of my system, and it will not have any problems. If you must say, you know that many people like to modify cars. Adding a third-party assisted driving system is actually the same as modifying a car. If there is a problem, the person who modified it will be responsible for it.

However, this cool-looking product is somewhat controversial in the industry.

Hou Xiaodi, CTO of Tucson Interconnect Technology, wrote in a column on Zhihu: George Hotz's driving solution, no matter how it is disguised and added with buzzwords such as Big Data and Deep Learning, has actually been determined in academia. In recent years, people are used to calling this approach End-to-End Learning, and earlier there were other aliases such as Imitation Learning. A major limitation of this type of method is that it can only handle input types that have been seen. If the situation encountered in actual testing is too different from the training data, it exceeds the upper limit of the model invariance.

● ● ●

The latest and most exciting content from GeekPwn can be found on Leiphone.com’s Home Channel. Scan the QR code to follow us.

In addition, this Wednesday at 14:00, Zhao Hanqing, a security researcher at Changting Technology and a member of the GeekPwn Hall of Fame, will share with us how to crack routers in the "Hardware Innovation Open Class". The number of participants is limited, so hurry up and scan the QR code to sign up!