The company’s Autopilot system is in active use, and in fact just got an update this past weekend that should vastly improve its performance and decrease the possibility that the car doesn’t brake when it should.
I am not gonna lie - that does not give me a lot of confidence in this system. "Oh, and by the way, it sometimes doesn't brake when it should." That quickly equals injury or even death with a car. You do not want to move fast and break things with your car.
How many crashes can be summed up as "the driver didn't break when they should"? The system doesn't have to be perfect, it just has to be better than the human. And actually, it doesn't even have to be better than the human: only the combination of the system and the human needs to be better than just the human alone.
All I can think about is the reliability of the hardware and how a failure could set back the market years (if not more due to regulation). I really hope this was properly engineered by seasoned automotive engineers.
My understanding is that this will only work with late-model Honda/Acura with LKAS & ACC packages, they hack into these systems for control. http://automobiles.honda.com/sensing/
OK so reading that, those new cars actually have most if not all of the self driving autopilot type features built in, very similar to Tesla (but not anywhere near Googles level). So I am guessing this device mainly just engages those built-in systems a little more frequently than they were designed? I mean what does the product actually do, or is it just basically Honda didn't want to market it as self driving yet but this guy decided he would?
> similar to Tesla (but not anywhere near Googles level).
With that comment you probably mean the inherent difference of the Tesla/Hotz/German style of live online interpretation of the world, compared to the inferior Google style offline collected Google maps, where the SW matches the view with the offline world. Because in the Google world you cannot drive a street which is not in the system. The Google system is not advanced enough to interpret changed speed signs, changed roads, temporary road contruction, ...
Is this product actually just limited to driving 280 / 101? If so, that's a really interesting way to launch a minimum viable product in the self-driving space: Obviously you can't launch with anything incomplete in terms of safety, but you can launch with a highly restricted supported geography.
He's saying that he's done that drive without touching the wheel. It will work in other places, but he emphasizes that you still have to pay attention, as with autopilot.
I guess you didn't watch the video. He goes out of his way to trash talk every big player in the industry. It's entertaining, informative, and cringeworthy.
I like George Hotz, but I was attempting to preempt anyone who might dismiss his substance because they don't like his style.
The only reason they can "launch" this is because the driver will still be held responsible for the actions of the car, no matter what their system does or doesn't do.
I.e, I can install a radar detector, but no way will I be able to sue the manufacturer if I get a ticket.
This isn't just a minimum viable product, it's a minimum liability product.
I think he just used that as an example of what types of journeys can be driven.
What mostly bothers me about this is his disregard to whether someone dies or not using the product. But hey, as long as they're shipping this year, under $1000, that's all that matters right?
I've read their paper and it's solid ML research, he's got a few good guys on the team. And I don't really disagree about MobilEye. But there are so many other DL/RL product spaces they could have picked without risking people's lives - it kind of seems he is on a mission to prove Musk and some other companies wrong, no matter what the risks to users. I sincerely hope no one dies as a result of using this.
It's silly to be throwing out dire warnings about Comma's product without first going after Tesla. With a level 2 system, whether it works 90% of the time, or 99.9% of the time, the driver has still got to watch the road and be ready to respond in an instant. In fact, the less reliable the system is, the less there is the likelihood that a driver will entrust their life to it when they shouldn't.
I agree - I think the way Musk has been so bullish in public about autonomous driving is dangerous. I think Tesla needs more sensors, in particular LIDAR, and better software too, and I'm not sure Autopilot is an apt name for a level 2 system like the one on Tesla or the new Volvos.
However, I think that difference between 90% and 99.99999% is crucial, and that boils down to the quality of the software and hardware, and the amount of testing. I find it irresponsible to beta test a product on users when there is a very real chance of something going wrong and people dying or getting injured.
What do you mean by "auto steering and braking"? Does this simply turn on settings that already exist in the car or does this actually control the car through fly by wire commands? If it is the latter, is this something that car manufacturers actively support third parties doing? To the layman like me, that seems like a huge potential liability.
The latter. I remember listening to an interview with George where he chose Honda/Acura because they're "generally hands off about these kind of things" meaning aftermarket mods.
Zipcar routinely starts and stops cars by satellite; OnStar pulls speeding fugitives over for the cops. In the 'this is bad' domain, some people say Michael Hastings was killed by remote software exploit of his car.
It is the same mechanism that allowed a Jeep's brakes to be hacked remotely, extended to the steering. So yes, car manufacturers do implement and ship extremely poorly thought-out ideas.
“It is fully functional. It’s about on par with Tesla
Autopilot,” Hotz said. It doesn’t have a lot of sensors
as the Comma One relies on built-in car front radars and
comes with a camera.
My car, a Passat 2016 has both of those and does an okay job keeping itself in the lane and such. Only with VW would be more aggressive about updating their cars firmware. They've got all the components to be pushing forward. But just like all the other car companies they'd rather wait until 2017 models to make any improvements.
Its probably built into their old school model that even if they could do firmware updates (likely still via USB), that they don't in order to get you to replace the entire vehicle sooner. Just speculating
I would love nothing more than for Bay Area congestion to be solved by self-driving cars, but I take this announcement with a grain of salt because it lacks any mention of liabilities and regulatory hurdles. Unless the management can answer the difficult questions, unforeseen difficulties likely await anyone who installs this device in their car. Insurance companies may refuse to cover accidents caused by the device, and automakers may even void the warranty if there is any demonstrable link between self-driving and increased engine wear & tear.
I am fairly confident that self-driving will proliferate a feature offered by the automaker at the time of purchase, just as adaptive cruise control and lane assist is becoming commonplace right now. It may be difficult to convince an average consumer to rig up a device in their car that isn't supported by their automaker.
That's a good point, although what I'm more concerned with is industry pushback against third-party self-driving upgrades, similar to how ECU tampering voids the warranty of most cars.
Autonomous software is a lot more complex than the software the ECU uses (which is still complex), so there may be a genuine argument against allowing people to modify cars not designed for autonomous driving. It definitely raises a really serious liability question.
Even with those systems working exactly as designed, there are some serious liability questions I don't think anyone has really answered yet.
It's a driver assistance feature, meant to be roughly on par with Tesla's Autopilot. There are no regulatory hurdles to clear here, as the tens of thousands of Teslas on the road can attest. Certainly there are huge regulatory and liability issues to address before full autonomy, but that's not what this is about.
>>I would love nothing more than for Bay Area congestion to be solved by self-driving cars
How would it do that though? I'm still going to take my car to work, except that it will do all the driving. Unless you are talking about theoretical, far off in the future idea that we will remove all traffic lights and speed limits because autonomous cars will handle intersections and speeds much better than humans?
This may be insensitive but is the name too close to "coma" given the odds that one of their users will end up in one (through the product's fault or not)?
Really now, Satan? That's all I had to do to get the top job in hell? Do I get to be the second coming for holding elevator doors when I'm in a rush?
I was just questioning the wisdom of passing out the free headline tragedy-puns with your branding. It would be an equally terribly name for a line of DIY transcranial direct current stimulation kits. Again, it's not because these products are dangerous[0], but that they will likely be perceived that way. Any possible blame will be laid at their feet at least through innuendo, so why make it easy and catchy?
[0] I make no assertion about the actual safety of either.
Huge props to Geohot for getting this to market (barring it ships). Curious - if it replaces the rear-view mirror, then what do you use as a rear-view mirror?
> The $999 price point is designed to be affordable, and is possible because of the components Comma uses in its product, which tend to be inexpensive off-the-shelf electronics.
"Inexpensive off-the-shelf electronics" aren't rated for automotive environmental conditions, nor do they have the immunity to interference (e.g., single event upsets) required for safety-critical systems.
Off-the-shelf in the context of electronics means that they didn't custom design some parts of it (for example using a SoC module rather than custom DDR layout), not that the parts are consumer rated.
So you could basically deploy this as software only, on recent 820 based smartphones? Just need a fish eye lens attachment and CAN bus to USB cable (or USB OTG GPIO)
It is more than just parts though. To get past the more demanding EM compatibility/immunity certs you have to design the whole product with that in mind, at which point it stops being off-the shelf or inexpensive.
Just wait until Comma.ai causes an accident and the NHTSA finds they can't do an analysis of the system because that system is an end-to-end neural network.
But if there was ever a case for a "business guy" to be involved with a pitch this was it.
His messaging works at some level - "I'm shipping something REAL and SMALL and ACHIEVABLE"
But his pitch has two huge problems:
(a) The delivery is rough. He comes across as patronizing and manic. I want to put my life in his hands?
(b) He fumbles (like Musk has) with the difference between his pitch "Mountain View to San Francisco without touching the wheel" and the reality "It is fully functional. It’s about on par with Tesla Autopilot"
Lots of potential in the product but for me this pitch diminished the chances that I'd ever own the device.
Yes, it's geohot's company but I don't think it lends it anymore credibility. Hacking is very different from building a viable product.
Moreover, some of his comments make me distrust his judgement. Just from a CNN article this year:
> "For awhile, I thought it was going to be incredibly difficult to contribute," he said. In January 2015, he started working as a researcher at Vicarious, a company that's building artificial intelligence algorithms. "I started reading all the papers. I thought, 'This stuff isn't that hard. This isn't ultra sophisticated. This is basic,'" he said. Moreover, "the mistakes people are making are basic."
> "I started to look around at the other players. These people are noobs," he said.
I don't think the target audience is the average consumer. I Think the target audience at the moment is extreme early adopter, SF types who think "Screw the system! I want to live in the future NOW!"
And this pitch certainly works for those type of people.
This is by far the sketchiest autonomous driving product I've seen so far.
Does the NHTSA do any testing for products like this? I'm usually not big on government regulation, but there definitely needs to be some kind of regulation and oversight for this kind of thing.
Wow, its incredible how many people bash this startup here, incredible to see it from this community here.
I mean what they have done to you? It's not like they gonna sell tickets to Hell or scam someone, so what's the big deal with more entrepreneurship in the town?? Be nice to your fellow entrepreneurship -- you might end up working for them one day.
I'm not saying their solution is perfect. Nothing is. But will I send a hate mail to utensil factory because I cut myself with their kitchen knife? No, I won't.
This is an extremely irresponsible attitude, and I'm shocked to see it expressed on HN!
As engineers / technologists / experts, we have a duty as a community to ensure our products are safe!
Of course, end users also have a duty to ensure they use our products responsibly -- but the whole point of autonomous vehicles is that the end user has no control. They are entrusting their safety to the product.
Do you think comma.ai is undervaluing it? Do you think they don't know that they will be under an irrational microscope the second they release an MVP?
Interesting that they're talking about California driving. In an episode of Bloomberg's Hello World I saw recently, Hotz actually demonstrated his self driving car in Nevada. They mentioned that the state of California asked him to stop driving on the public roads.
Surprised with all the skepticism here. It's not perfect but this guy is a hardware, software, and AI genius and this could very well speed up the self driving race by several years or even more. When it comes to price, it could speed things up by more than a decade. With some funding and a solid team this could move incredibly quickly.
Some thoughts in no particular order, edited from an IRC discussion I just had on this:
The price is extremely low. If you drive just 20 minutes per day, and your time is worth at least minimum wage, it makes sense to buy this for just a year, and hopefully it lasts longer than that. But there is a monthly subscription cost ($300 per year) i didn't factor in. And wtf, since when does hardware have a subscription cost? Oh, and to purchase one you need to have lots of "points" with the company, which you can earn by using their app. This is really weird, why not just raise the price if you can't meet demand?
The price also doesn't seem high enough to cover the risk. If each accident costs 10 million in lawsuits (i have no idea how much they will actually cost, but that seems like a good made up number, and corporate legal fees are expensive anyway) then they better have less than 1 accident per 10,000 vehicles that use it, just to break even. honestly i hope they have 0 accidents because they could bring in regulators and ruin the fun for everyone, as well as the very delicate reputation of self driving cars (people are already predisposed to distrust machines)
I can't believe this works at all. Wouldn't you need like multiple gpus to process all this camera data in real time? How do you account for the variance between vehicles? How many cameras do they have and what is the resolution? That unit looks really small.
It looks like it's just trained by predicting what a human would do. Human drivers are imperfect, and so I imagine this machine will learn to imitate the imperfections of human drivers, combined with it's own imperfections. It's also unclear how they trained it, because the dash cam videos don't give information on when users tapped the brakes or moved the steering wheel. They must be doing some interesting semi-supervised learning.
Apparently the cars have limits on how hard it can slam on the brakes or turn the wheel. He advertises this as a good thing that will prevent the AI from doing anything too bad, but I don't see how. This limits what the AI can do in a real emergency. Worse, it suggests that the AI needs these training wheels. That without them it would cause accidents, and that's very concerning.
I just can't believe that this works. Deep learning is completely amazing and can get 99% accuracy, but this requires 99.9999%. Those extra 9's are super super hard. Fortunately this doesn't need to be perfect, it only needs to be better than humans. And while deep learning is starting to beat humans in some machine vision tasks, I'm not confident this is true in general. In the real world there will be weird edge cases, like a car painted in a very atypical pattern, or ice on the road, or a bug on the windshield, etc. Humans are more capable of handling new, out of sample situations, and machines can fail epically at them.
Yes but that's very different. Most accidents aren't the suing a company but a driver. This would be more like someone suing a car manufacturer because they made a faulty steering wheel or something. And I based that on the economic value of life, as estimated by various government agencies to be around that number: https://en.wikipedia.org/wiki/Value_of_life#Life_Value_in_th... But it should also include legal fees which aren't cheap at all.
You're also assuming fatalities which are not common. You're also assuming that the software company will be able to be found liable after both a waiver and a proven history of safety testing. Currently we have about 15 fatalities per 100k drivers. 40% of those are alcohol related so I'm going to count those as saves right away. Many of the rest are weather/sleeping/texting/etc that are easily savable. Even if we pretend those will all still happen, the goal is a 10x safety increase which would mean .9 per 100k. Even at your outlandish 10M number that would not kill the company, but the reality is closer to 1/10 that amount and even less if the legal hurdles are properly-managed.
> And wtf, since when does hardware have a subscription cost
Well there is a pretty important software component and I'd guess that this allows them to fund ongoing development of that and give you regular updates (this is a pretty early version so I'd imagine at least monthly software updates after purchase).
> Wouldn't you need like multiple gpus to process all this camera data in real time
Hotz demoed an early prototype on the road in Nevada to a reporter months ago that was wired up to a single Ubuntu-Linux powered laptop if I remember correctly so I'm assuming probably just 1 integrated GPU required.
They say they get the video feed. I can imagine why they want it (training?) but having somebody potentially watching what's in front of your car all the time is much worse than having them know where you are. I won't buy that product. However I think that if this really works it will sell well.
It may well be the strategy - I'm not convinced they actually expect people to stop worrying about death, buy and use the $999 box. They will certainly not ship large amounts of units by end of the year, and risk several accidents at this stage with effectively very alpha software/hardware combo.
I think they will get somewhere, if not in self driving cars then something else to do with machine learning.
This guy is a different breed. He had a chance to work for Tesla, but he wanted to do it himself. I hope I'm not just in the fog on this, but I want to believe that he is a true disruptor.
115 comments
[ 2.5 ms ] story [ 197 ms ] threadI am not gonna lie - that does not give me a lot of confidence in this system. "Oh, and by the way, it sometimes doesn't brake when it should." That quickly equals injury or even death with a car. You do not want to move fast and break things with your car.
fixed typo
With that comment you probably mean the inherent difference of the Tesla/Hotz/German style of live online interpretation of the world, compared to the inferior Google style offline collected Google maps, where the SW matches the view with the offline world. Because in the Google world you cannot drive a street which is not in the system. The Google system is not advanced enough to interpret changed speed signs, changed roads, temporary road contruction, ...
He gives the full talk here:
https://techcrunch.com/video/george-geohot-hotz-presents-the...
Yes, he's a punk and seemingly on drugs, but surprisingly, Hotz is not a bullshitter as far as I can tell.
What makes you say this?
I like George Hotz, but I was attempting to preempt anyone who might dismiss his substance because they don't like his style.
The dangerous kind of not-really-autonomous.
I.e, I can install a radar detector, but no way will I be able to sue the manufacturer if I get a ticket.
This isn't just a minimum viable product, it's a minimum liability product.
If you start tampering with the hardware of the car to splice in this box, will your insurance cover even in normal situations?
What mostly bothers me about this is his disregard to whether someone dies or not using the product. But hey, as long as they're shipping this year, under $1000, that's all that matters right?
I've read their paper and it's solid ML research, he's got a few good guys on the team. And I don't really disagree about MobilEye. But there are so many other DL/RL product spaces they could have picked without risking people's lives - it kind of seems he is on a mission to prove Musk and some other companies wrong, no matter what the risks to users. I sincerely hope no one dies as a result of using this.
However, I think that difference between 90% and 99.99999% is crucial, and that boils down to the quality of the software and hardware, and the amount of testing. I find it irresponsible to beta test a product on users when there is a very real chance of something going wrong and people dying or getting injured.
https://electrek.co/2016/09/14/another-fatal-tesla-autopilot...
I am fairly confident that self-driving will proliferate a feature offered by the automaker at the time of purchase, just as adaptive cruise control and lane assist is becoming commonplace right now. It may be difficult to convince an average consumer to rig up a device in their car that isn't supported by their automaker.
Even with those systems working exactly as designed, there are some serious liability questions I don't think anyone has really answered yet.
How would it do that though? I'm still going to take my car to work, except that it will do all the driving. Unless you are talking about theoretical, far off in the future idea that we will remove all traffic lights and speed limits because autonomous cars will handle intersections and speeds much better than humans?
I was just questioning the wisdom of passing out the free headline tragedy-puns with your branding. It would be an equally terribly name for a line of DIY transcranial direct current stimulation kits. Again, it's not because these products are dangerous[0], but that they will likely be perceived that way. Any possible blame will be laid at their feet at least through innuendo, so why make it easy and catchy?
[0] I make no assertion about the actual safety of either.
"Inexpensive off-the-shelf electronics" aren't rated for automotive environmental conditions, nor do they have the immunity to interference (e.g., single event upsets) required for safety-critical systems.
So you could basically deploy this as software only, on recent 820 based smartphones? Just need a fish eye lens attachment and CAN bus to USB cable (or USB OTG GPIO)
But if there was ever a case for a "business guy" to be involved with a pitch this was it.
His messaging works at some level - "I'm shipping something REAL and SMALL and ACHIEVABLE"
But his pitch has two huge problems:
(a) The delivery is rough. He comes across as patronizing and manic. I want to put my life in his hands?
(b) He fumbles (like Musk has) with the difference between his pitch "Mountain View to San Francisco without touching the wheel" and the reality "It is fully functional. It’s about on par with Tesla Autopilot"
Lots of potential in the product but for me this pitch diminished the chances that I'd ever own the device.
Wait, Hotz, that name... is this Geohot?
...
Turns out, it is! [1] That makes this statement a lot more believable to me.
[1] https://en.wikipedia.org/wiki/George_Hotz
Moreover, some of his comments make me distrust his judgement. Just from a CNN article this year:
> "For awhile, I thought it was going to be incredibly difficult to contribute," he said. In January 2015, he started working as a researcher at Vicarious, a company that's building artificial intelligence algorithms. "I started reading all the papers. I thought, 'This stuff isn't that hard. This isn't ultra sophisticated. This is basic,'" he said. Moreover, "the mistakes people are making are basic."
> "I started to look around at the other players. These people are noobs," he said.
[1] http://money.cnn.com/2016/04/04/technology/george-hotz-comma...
And this pitch certainly works for those type of people.
Does the NHTSA do any testing for products like this? I'm usually not big on government regulation, but there definitely needs to be some kind of regulation and oversight for this kind of thing.
I mean what they have done to you? It's not like they gonna sell tickets to Hell or scam someone, so what's the big deal with more entrepreneurship in the town?? Be nice to your fellow entrepreneurship -- you might end up working for them one day.
Apparent reckless use of technology frightens people, regardless of the actual risk.
As engineers / technologists / experts, we have a duty as a community to ensure our products are safe!
Of course, end users also have a duty to ensure they use our products responsibly -- but the whole point of autonomous vehicles is that the end user has no control. They are entrusting their safety to the product.
The situations are in no way comparable.
That difference is engineering, and we undervalue it at our peril.
What happens when the knife cuts you? :-)
Interesting that they're talking about California driving. In an episode of Bloomberg's Hello World I saw recently, Hotz actually demonstrated his self driving car in Nevada. They mentioned that the state of California asked him to stop driving on the public roads.
Regulations ahoy!
The price is extremely low. If you drive just 20 minutes per day, and your time is worth at least minimum wage, it makes sense to buy this for just a year, and hopefully it lasts longer than that. But there is a monthly subscription cost ($300 per year) i didn't factor in. And wtf, since when does hardware have a subscription cost? Oh, and to purchase one you need to have lots of "points" with the company, which you can earn by using their app. This is really weird, why not just raise the price if you can't meet demand?
The price also doesn't seem high enough to cover the risk. If each accident costs 10 million in lawsuits (i have no idea how much they will actually cost, but that seems like a good made up number, and corporate legal fees are expensive anyway) then they better have less than 1 accident per 10,000 vehicles that use it, just to break even. honestly i hope they have 0 accidents because they could bring in regulators and ruin the fun for everyone, as well as the very delicate reputation of self driving cars (people are already predisposed to distrust machines)
I can't believe this works at all. Wouldn't you need like multiple gpus to process all this camera data in real time? How do you account for the variance between vehicles? How many cameras do they have and what is the resolution? That unit looks really small.
It looks like it's just trained by predicting what a human would do. Human drivers are imperfect, and so I imagine this machine will learn to imitate the imperfections of human drivers, combined with it's own imperfections. It's also unclear how they trained it, because the dash cam videos don't give information on when users tapped the brakes or moved the steering wheel. They must be doing some interesting semi-supervised learning.
Apparently the cars have limits on how hard it can slam on the brakes or turn the wheel. He advertises this as a good thing that will prevent the AI from doing anything too bad, but I don't see how. This limits what the AI can do in a real emergency. Worse, it suggests that the AI needs these training wheels. That without them it would cause accidents, and that's very concerning.
I just can't believe that this works. Deep learning is completely amazing and can get 99% accuracy, but this requires 99.9999%. Those extra 9's are super super hard. Fortunately this doesn't need to be perfect, it only needs to be better than humans. And while deep learning is starting to beat humans in some machine vision tasks, I'm not confident this is true in general. In the real world there will be weird edge cases, like a car painted in a very atypical pattern, or ice on the road, or a bug on the windshield, etc. Humans are more capable of handling new, out of sample situations, and machines can fail epically at them.
Well there is a pretty important software component and I'd guess that this allows them to fund ongoing development of that and give you regular updates (this is a pretty early version so I'd imagine at least monthly software updates after purchase).
> Wouldn't you need like multiple gpus to process all this camera data in real time
Hotz demoed an early prototype on the road in Nevada to a reporter months ago that was wired up to a single Ubuntu-Linux powered laptop if I remember correctly so I'm assuming probably just 1 integrated GPU required.
Hopefully sane legislation will prevail and end this madness soon.
Also avoiding liability.
I think they will get somewhere, if not in self driving cars then something else to do with machine learning.