Tesla Chip-Maker NVIDIA Demonstrates Self-Driving Car That Uses AI

OCT 16 2016 BY MARK KANE 10

Inside the Tesla Model S, Image Credit: Tesla

Inside the Tesla Model S, Image Credit: Tesla

NVIDIA, which is Tesla Motors’ supplier for the Visual Computing Modules (VCM), is also working on autonomous driving technology.

NVIDIA’s approach is different from conventional methods and relies someewhat on artificial intelligence.

The company’s DRIVE PX 2 AI car computing platform learns how to deal with various situations on the road as it experiences them.

As Tesla abandons MobileEye hardware, NVIDIA is hinted as a possible new supplier for new generation Autopilot.

“In contrast to the usual approach to operating self-driving cars, we did not program any explicit object detection, mapping, path planning or control components into this car. Instead, the car learns on its own to create all necessary internal representations necessary to steer, simply by observing human drivers.

The car successfully navigates the construction site while freeing us from creating specialized detectors for cones or other objects present at the site. Similarly, the car can drive on the road that is overgrown with grass and bushes without the need to create a vegetation detection system. All it takes is about twenty example runs driven by humans at different times of the day. Learning to drive in these complex environments demonstrates new capabilities of deep neural networks.

The car also learns to generalize its driving behavior. This video includes a clip that shows a car that was trained only on California roads successfully driving itself in New Jersey.”

The CEOs of both companies have already talked about autonomous driving in an interview with NVIDIA CEO Jen-Hsun Huang:

source: Teslarati

Categories: General, Tesla

Tags: ,

Leave a Reply

10 Comments on "Tesla Chip-Maker NVIDIA Demonstrates Self-Driving Car That Uses AI"

newest oldest most voted
This sounds like the approach that I expected computer hardware and software engineers and developers to use, rather than what Tesla is talking about… which is to build up an enormous detailed library of individual stationary obstacles, and the exact configuration of roads, in areas where Tesla cars are being driven. Seems to me that such an approach — which seems to be a heuristic learning programming approach — would be far better and far more flexible. Such flexibility will be needed in the real world, where construction zones and detours may pop up unexpectedly on any day. It seems to me that such flexibility would also be far preferable in such cases as, for example, driving down the freeway and encountering a mattress lying in the road, one that has fallen off an open bed truck being used as a moving van. That’s a real-world example of something that happened to us not so long ago. However, I’m certainly willing to be convinced I’m wrong about this, if the evidence shows otherwise. It’s good that different companies are taking different approaches to developing self-driving cars. I think it will become apparent over the next very few years just what… Read more »

I think the purpose of highly detailed mapping is to let the AI quickly filter out what’s “supposed” to be there and focus all its attention on the other stuff. Plus to provide context to help interpret the other stuff.

I expect custom silicon from Tesla for the next AutoPilot. They’ve hired a number of CPU designers and execs from AMD…

News about hiring AMD people is kinda strange.

AMD is open to do semi custom jobs (like for Sony and Microsoft).

Should be cheaper for Tesla to outsource (at least first iteration) of that hw to AMD who have people and hardware to create chips.
(Chips simulators costs upwards a milion bucks)

Those chip engineers are probably designing small specific stuff, and nothing like the main SOC’s.

These were all pretty low speed tests which is where, as one would expect, this sort driving works best.
You have to crawl before you can walk. So at low speeds the computer has more time to react to conditions and do the right thing, though at higher speeds that ability will diminish, not being as accurate, since the optimum action must occur within a shorter time window.
In the long run some sort of hybrid system which incorporates a number of approaches will evolve.

I agree wholeheartedly with this approach, partly because it widens the autonomous sphere outside heavily marked, kerbed, signed, standardized roads in the top of the developed world. Being able to deal with potholes and the untidy, crazy stuff in the majority of the world is what will make this global.

Sorry but no.

Mobileeye was NOT a provider of self driving solution to Tesla. Tesla did it’s own autopilot.

“The company’s DRIVE PX 2 AI car computing platform learns how to deal with various situations on the road as it experiences them.”

Isn’t this the exact same approach that acclaimed hacker/programmer George Hotz aka Geohot is using to create the $999n Comma One autonomous driving add-on. And didn’t Elon try to hire Hotz to help develop Autopilot at Tesla using this “self learning” approach?

http://insideevs.com/public-split-between-tesla-and-autopilot-chip-provider-mobileye-gets-messy/

That’s the way most researchers approach it. Google, too. Car companies are more of a mixed bag. Some see it as an extension to auto-braking and lane-keeping, which tend to be rule-based.

One problem with learning-based approaches is it’s difficult to nail down a specific “flaw” when something goes wrong. The NTSB wants you to prove you found and fixed the problem.