Slashdot Mirror
Hacker George Hotz Unveils $999 Self-Driving Add-On (pcmag.com)
An anonymous reader quotes a report from PC Magazine: Hacker George Hotz is gearing up to launch his automotive AI start-up's first official product. In December, the 26-year-old -- known for infiltrating Apple's iPhone and Sony's PlayStation 3J -- moved on to bigger things: turning a 2016 Acura ILX into an autonomous vehicle. According to Bloomberg, Hotz outfitted the car with a laser-based radar (lidar) system, a camera, a 21.5-inch screen, a "tangle of electronics," and a joystick attached to a wooden board. Nine months later, the famed hacker this week unveiled the Comma One. As described by TechCrunch, the $999 add-on comes with a $24 monthly subscription fee for software that can pilot a car for miles without a driver touching the wheel, brake, or gas. But unlike systems currently under development by Google, Tesla, and nearly every major vehicle manufacturer, Comma.ai's "shippable" Comma One does not require users to buy a new car. "It's fully functional. It's about on par with Tesla Autopilot," Hotz said during this week's TechCrunch Disrupt in San Francisco.
Does he indemnify the users in case his tech screws up and kills someone?
Going cheap can get very, very, VERY expensive, very quickly.
No, it's not a shame because his system is very dangerous. He built it using reinforcement learning based off his personal driving patterns. Meaning he drove around in training mode and it watched what he did. Then in running mode it copies what he did in similar situations. So that means you can't count on it in any new or lesser seen situations. Any event that didn't occur often enough during training, like a bird flying infront of you, can cause unexpected behavior by the AI. This type of system is inherently unsafe unless you train it against everything. You need to train both what to do and what not to do.
How often has he had someone swerve into his lean or cross in front of him? How many times did someone almost do that? Unless you specifically train for when that happens, this type of learning system could learn to see the natural drifts of people driving between the lines as a hard rule and then freak out when someone suddenly or slowly mergers in front of you. If most pedestrians it sees walk up to the edge of the road and then wait the car passes, it'll never learn to slow down for them because it never saw past that waiting point. Thus when someone does step into the street, you'll run them over before you realize the AI isn't going to stop for them.
There was a research paper who's algorithm made a massive improvement in identifying cars from photographs. Well they thought it did until other people looked closer and figured out all the images with cars were taken later in the day and all the non-car pictures were taken earlier in the day. The AI was only checking the sky color and using that to determine if a car was in the picture or not. Excellent in the training result, horrible real-life performance. The way he trained his system allows this same type of issue to occur.
I have an '07 S80 V8 with adaptive cruise and it's my favorite feature. 2007 was kind of early in the adaptive cruise era but it works well, down to 20 mph but won't do a complete stop, though the collision avoidance system will pre-charge the brake pressure.
The problem with adding the hardware, even with a modern Volvo that offered the feature in the same model as an option is that almost nothing in a Volvo can be added without VIDA installing software. My S80 spent 4 days in the shop when I added the iPod connector because they had to load a custom software patch to make the component work.
I think adaptive cruise would be about the sweet spot as a "tack on" -- simple enough to patch it into the existing cruise control system. But lane control and more advanced autonomous stuff would be a lot of mechanical additions, not counting software or electronics even.