> Whether they are public transit agencies or the Ubers and Lyfts of the world or logistics companies like Amazon or FedEx or UPS. We think about building a driver that can support all of those opportunities.
Why not just focus on one of these? Or even just highway driving kits first for long hauls and commuters?
This seems to be the approach geohot is trying to take with his company. Narrowing the scope down to the simple stuff and working it into serious commercial operations, then scaling it up to harder stuff.
The main thing is scope creep with the different demands of the different drivers.
But who knows maybe they have a different internal plan for initial roll out. Or a more generic use case is indeed the best approach...
This is the kind of thing you tell investors to justify a higher valuation because the total addressable market is potentially massive. I wouldn't read it too literally. They're very likely to be focusing on just 1-2 of these to start.
On the other hand...the road is the road.
It's not like the logistics companies are trying to automate train traffic, public transit is all about them boats and others are into the road business.
They're all on the road and the main thing that might differ is the behaviour of their transport to a certain extent (eg trucks vs cars)
You could focus on highways only for logistics transport but even then you might need to be able to react to a sudden pedestrian...even if it's just to make truckers lives a bit easier because you can always rely on human error.
B2B logisitcs have fixed routes so one can simplify a lot of things. For example, you can map out all traffic lights, stop signs, turns, lanes precisely and without ambiguity. This can reduce reliance on stochastic predictions by an order of magnitude. Also, fixed routes mean continuous recurring paid high value business.
Not really though, right? I theoretically have the same commute every day, but there have been times when a moving truck has been stopped in my lane, or the period when a road was down to one lane as it was repaved, and then re-opened with a new traffic circle, etc. You can't train your truck how to complete a route today and then assume it'll be able to successfully complete that route everyday as detours, new traffic patterns, accidents, etc, happen.
Interstates are more consistent than local roads. I'm surprised long haul trucking isn't the first target for these companies. Driving 20 hours straight and pulling up next to a local driver for the last mile would be faster (human drivers are required to sleep) and save a bunch of money.
Consistency is key in automation.
Other than the cost of automation itself consider the big big costs when those irregularities do exist, the big cost of the systems made to account for and handle those potential failures and the fact that as almost always there's more of those failure points than you expect.
All of these things are partnerships (to one degree or another) of which SDCs are a piece of the larger puzzle of operations, logistics, regulations, etc.
I agree with you that scoping down to one scenario makes sense, but they'd need to find a highly motivated partner to do that with.
The article did address that apparently the necessary sensors for commercial trucking use cases have only recently become available. I'm guessing it's related to how far the systems can look? Due to the involved speeds and stopping distances, truck systems need to be looking very far ahead. Anything nearby and the truck needs to be maneuvering instead of stopping, which is also non-trivial with the trailer dynamics.
It's Urmson, who likes conservative, safe designs. They may be the winner in the end.
The competition in self-driving is thinning. Uber gave up. Volvo gave up. Apple gave up. Tesla is stuck at the driver-assistance level. Cruise is having problems, but they do have lots of test vehicles on the road. Waymo is soldiering on, but their plug might be pulled. Google/Alphabet are cutting back on "moonshots", not having had much success with any of them. (See list from 2016. Any successes?)[1]
Are you sure Apple actually gave up yet? I thought they still have related activity following their recent acquisition of Drive.ai.
If anyone, it could well be Apple that can catch up with Tesla's self-driving capabilities.
When it comes to Tesla, they might be sitting on better technology than it might seem at first glance.
Although I'm not sure they'll manage in Europe, pedestrians and cyclists are more common and have better rights. Does Tesla get enough non-visual wavelength information from front sides, to avoid pedestrians and wild animals? Tesla's self-driving story could be cut short in Europe the first time the car kills a cyclist crossing the road in low visibility conditions, like during heavy snowfall.
Can Tesla manage in low angle sunshine and snow or dirt on the cameras? Humans can shift their viewpoint, but cameras can't. This could be particularly difficult in North Europe.
Some Asian markets like India and Vietnam could be difficult for them because of chaotic traffic systems.
Addendum: If played "right", self-driving can be $100B+ business. I'd expect true poker hands to be only shown once time is right and legislation allows it.
Having used Siri, my optimism about Apple AI and technology, above all other competitors, powering a car are not great. It’s significantly behind Google and Amazon on that front. Wouldn’t it be odd if they were suddenly the world leader in self-driving vehicles?
No. There is very little overlap in technology and definitely no overlap in the teams at all.
Google’s instant messaging landscape is shit, but that doesn’t mean their email isn’t decent. And those two products have more overlap than self-driving cars and a voice assistant.
>Are you sure Apple actually gave up yet? I thought they still have related activity following their recent acquisition of Drive.ai.
Back when I was working in the field of VR, we were all fully convinced that apple was about to go enter the field with some huge breakthrough, since pretty much every niche product for vr devs was being bought as discretely as possible. It was really common to go to some website to find their software or manuals and find out it redirected to apple.com with no further explanation.
Five years have passed since then with nothing to show up for... So I guess the lesson is that you shouldn't take investments as a sign of something cooking, at least not for giants like apple that can afford to spend millions "just in case"
> I guess the lesson is that you shouldn't take investments as a sign of something cooking
Leaving aside acquihires, I think it's always a sign of something cooking. Apple just seems atypically content to let things simmer until they feel they've got a compelling product, and they aren't afraid to kill things that don't work.
Sometimes it happens quickly; Fingerworks was acquired in 2005 and the iPhone was released in 2007. The AuthenTec acquisition closed in 2012, and Touch ID showed up in 2013.
On the other side, you have acquisitions like PrimeSense. Apple bought them in 2013, and Face ID didn't show up until 2017.
That's Tesla's usual whine. Uber had the same complaint until they ran down a pedestrian. California's DMV is willing to license autonomous vehicles right now.[1] The manufacturer have to demonstrate that they work safely, and accept full responsibility for anything that goes wrong, of course. Some wannabe self-driving car companies have a problem with that.
This has nothing to do with attention. If the cameras' view is obscured by something, no amount of computing power will change that, unless some physical action can be taken.
Most of the time a human can just shift head a bit to see past whatever is blocking the view.
You do it all the time without even realizing. For example to see your phone's screen clearly when a reflection obstructs your view. Or to see behind whatever happens to be blocking your line of sight.
The possible solutions for this are obvious: add more forward-facing cameras and augment with something that's not affected (for example radar, lidar, infrared, etc.). Adding some physical feature that gives the camera some degrees of freedom to move could work as well, although it does add expense and lowers reliability.
I live in Cupertino and I see all their iterations driving up and down the street all the time (saw one last week). In fact, after the driveai aquihire, they changed their entire design to look like a mattress on top of a car. They’re still in it, maybe for Apple Maps or for something else.
Replace Amazon with AWS, and it helps at least to categorize AWS Aurora as some kind of IT SaaS. That Aurora is a DB-as-a-Service thing. They offer MySQL and PostgreSQL.
How about instead of using lidar, we use a camera on a swivel that can check mirrors and look out onto the road with basically the same visual sight as a human. Then for test data we just have our drivers wear some kind of camera glasses to capture what they are looking at and have them drive around, create labelled data in terms of what they see and how they respond in terms of input to the car. Then we can train a system that just kind of drives like humans do, with the same flaws, but hopefully with the same strengths as well.
The advantage of lidar is not swivelling, but that it directly provides distance information (whereas with cameras you need to extract that from the parallax of two images).
It's not clear to me what advantage a swivelling camera would bring; Tesla just has cameras for all directions, which is likely cheaper and more reliable.
hm yes I guess I misspoke when I said swivelling camera. But basically I was thinking of providing the same sorts of visual inputs as a human receives, and then training on actual human drivers driving.
Because people aren't going to trust self-driving cars unless they're drastically better than humans are. Being as good as a human will always lead to an inevitable crash where the human will say "I could have avoided that" and now the whole system crumbles.
Human pedestrians get hit and killed by human-driven vehicles every day. Uber's self driving vehicle did it once, in a situation that was also missed by the human supervisor, and it was national news for weeks.
> Because people aren't going to trust self-driving cars unless they're drastically better than humans are.
I believe that people aren't ever going to trust self-driving cars, unless they are "perfect". Even ten 9s of reliability (no driving errors or accidents) would probably be too imperfect for people to accept (and I don't believe humans have ever made anything that close to perfection).
I believe the reason this is probably so has to do with assignment of blame and guilt. Something easily done with human drivers, less so (or maybe impossible) for self-driving vehicles.
Eventually I think they will, as the generations who drove their own car age out and new generations who grew up reliant on self-driving technology comes of age.
You see it with automatic transmissions, some people don't trust the computer to shift gears for them because they think they can do it better. As time went on and automatic gearboxes became more reliable and more popular, those complaints slowed down. (speaking from an American standpoint here)
You see it with anti-lock brakes (ABS). My grandparents still don't trust ABS even with studies showing ABS stops quicker than humans pumping the brakes, because it takes a level of control away from them. But that's not a common complaint these days.
You see it with electric cars, where any fire make national news but no one seems to care about the thousands of ICE cars that burn to the ground every year.
You even see it with seatbelts, people who claim it's safer to be thrown from the car in an accident.
But eventually those complaints lose their novelty and people just accept it as normal.
Yes you may be right about what it will take to win public trust. But if I'm not mistaken self-driving cars haven't yet reached the ability of human drivers. If we were to just deploy them willy-nilly the incidents of bad-driving/accidents they cause would go up, no?
There's a reason why almost everyone pursuing autonomous driving uses lidar despite the high costs. It's because lidar has extremely good "recall" and "precision" in most driving conditions. In other words, if an object is present on or near the road, lidar will almost certainly detect it (high recall), and it has very rare false positive detections (high precision). Lidar also directly provides distance information. For pure camera based approaches, we've seen huge improvements in recall in the past decade, but unfortunately we still need orders of magnitude reductions in the false negative rate to be good enough for safety critical applications. I think pure computer vision approaches will get there someday, but we first need some fundamental changes in CV algorithms in my opinion. Hopefully with increasing adoption, lidar will get significantly cheaper.
> Back in 2015, Urmson said that his goal for autonomous vehicles was that things would progress so rapidly that his 11-year old son would not need to get a driver’s license. Now, he has a more measured prediction: over the next five years, we’ll start to see commercial fleets of autonomous vehicles piloting people and goods. After that, we could start to see broad adoption.
So five years ago, he was sure we were five years away. Now he's sure we are five years away. How is this a "more measured prediction?" Because instead of completely generalized self-driving used by everyone, he's predicting... completely generalized self-driving used only by companies?
I predict that five years from now, Umson will have a more measured prediction that is still somehow five years away.
51 comments
[ 6.3 ms ] story [ 127 ms ] threadWell they just did that, didn’t they?
> that finally appears to be changing
Why not just focus on one of these? Or even just highway driving kits first for long hauls and commuters?
This seems to be the approach geohot is trying to take with his company. Narrowing the scope down to the simple stuff and working it into serious commercial operations, then scaling it up to harder stuff.
The main thing is scope creep with the different demands of the different drivers.
But who knows maybe they have a different internal plan for initial roll out. Or a more generic use case is indeed the best approach...
You could focus on highways only for logistics transport but even then you might need to be able to react to a sudden pedestrian...even if it's just to make truckers lives a bit easier because you can always rely on human error.
I agree with you that scoping down to one scenario makes sense, but they'd need to find a highly motivated partner to do that with.
The competition in self-driving is thinning. Uber gave up. Volvo gave up. Apple gave up. Tesla is stuck at the driver-assistance level. Cruise is having problems, but they do have lots of test vehicles on the road. Waymo is soldiering on, but their plug might be pulled. Google/Alphabet are cutting back on "moonshots", not having had much success with any of them. (See list from 2016. Any successes?)[1]
[1] https://www.businessinsider.com/20-moonshot-projects-by-goog...
If anyone, it could well be Apple that can catch up with Tesla's self-driving capabilities.
When it comes to Tesla, they might be sitting on better technology than it might seem at first glance.
Although I'm not sure they'll manage in Europe, pedestrians and cyclists are more common and have better rights. Does Tesla get enough non-visual wavelength information from front sides, to avoid pedestrians and wild animals? Tesla's self-driving story could be cut short in Europe the first time the car kills a cyclist crossing the road in low visibility conditions, like during heavy snowfall.
Can Tesla manage in low angle sunshine and snow or dirt on the cameras? Humans can shift their viewpoint, but cameras can't. This could be particularly difficult in North Europe.
Some Asian markets like India and Vietnam could be difficult for them because of chaotic traffic systems.
Addendum: If played "right", self-driving can be $100B+ business. I'd expect true poker hands to be only shown once time is right and legislation allows it.
Google’s instant messaging landscape is shit, but that doesn’t mean their email isn’t decent. And those two products have more overlap than self-driving cars and a voice assistant.
Seems easy enough to automate
Back when I was working in the field of VR, we were all fully convinced that apple was about to go enter the field with some huge breakthrough, since pretty much every niche product for vr devs was being bought as discretely as possible. It was really common to go to some website to find their software or manuals and find out it redirected to apple.com with no further explanation.
Five years have passed since then with nothing to show up for... So I guess the lesson is that you shouldn't take investments as a sign of something cooking, at least not for giants like apple that can afford to spend millions "just in case"
Leaving aside acquihires, I think it's always a sign of something cooking. Apple just seems atypically content to let things simmer until they feel they've got a compelling product, and they aren't afraid to kill things that don't work.
Sometimes it happens quickly; Fingerworks was acquired in 2005 and the iPhone was released in 2007. The AuthenTec acquisition closed in 2012, and Touch ID showed up in 2013.
On the other side, you have acquisitions like PrimeSense. Apple bought them in 2013, and Face ID didn't show up until 2017.
So might need to be a little more patient given that self driving cars is quite a bit more involved.
That's Tesla's usual whine. Uber had the same complaint until they ran down a pedestrian. California's DMV is willing to license autonomous vehicles right now.[1] The manufacturer have to demonstrate that they work safely, and accept full responsibility for anything that goes wrong, of course. Some wannabe self-driving car companies have a problem with that.
[1] https://www.dmv.ca.gov/portal/wcm/connect/a6ea01e0-072f-4f93...
Humans have to shift their viewpoint; computers need not.
I don’t think there’s agreement on what “attention” is, but I also think there is agreement that it exists.
Computers, on the other hand, can be 100% attentive to as many locations as we want them to be (limited by costs, energy use, etc)
Most of the time a human can just shift head a bit to see past whatever is blocking the view.
You do it all the time without even realizing. For example to see your phone's screen clearly when a reflection obstructs your view. Or to see behind whatever happens to be blocking your line of sight.
The possible solutions for this are obvious: add more forward-facing cameras and augment with something that's not affected (for example radar, lidar, infrared, etc.). Adding some physical feature that gives the camera some degrees of freedom to move could work as well, although it does add expense and lowers reliability.
Wait what? Tesla is not even in the top 3 of the autonomous pack. Why would Apple be trying to catch up to a distant 4th or 5th?
Tesla is leading in marketing though, if that helps :)
https://five.ai/faq
https://wayve.ai/
I live in Cupertino and I see all their iterations driving up and down the street all the time (saw one last week). In fact, after the driveai aquihire, they changed their entire design to look like a mattress on top of a car. They’re still in it, maybe for Apple Maps or for something else.
https://thelastdriverlicenseholder.com/2019/08/06/apple-car-...
I have some pics of my own if anyone's interested :)
https://twitter.com/suryadsuryad/status/1184912385624010752?...
https://twitter.com/suryadsuryad/status/1215836960603852800?...
I can not guess from the name "Aurora" what this is about,
and certainly not from the cryptic: "... is finally ready to show the world what it’s been up to"
It's not clear to me what advantage a swivelling camera would bring; Tesla just has cameras for all directions, which is likely cheaper and more reliable.
Human pedestrians get hit and killed by human-driven vehicles every day. Uber's self driving vehicle did it once, in a situation that was also missed by the human supervisor, and it was national news for weeks.
I believe that people aren't ever going to trust self-driving cars, unless they are "perfect". Even ten 9s of reliability (no driving errors or accidents) would probably be too imperfect for people to accept (and I don't believe humans have ever made anything that close to perfection).
I believe the reason this is probably so has to do with assignment of blame and guilt. Something easily done with human drivers, less so (or maybe impossible) for self-driving vehicles.
You see it with automatic transmissions, some people don't trust the computer to shift gears for them because they think they can do it better. As time went on and automatic gearboxes became more reliable and more popular, those complaints slowed down. (speaking from an American standpoint here)
You see it with anti-lock brakes (ABS). My grandparents still don't trust ABS even with studies showing ABS stops quicker than humans pumping the brakes, because it takes a level of control away from them. But that's not a common complaint these days.
You see it with electric cars, where any fire make national news but no one seems to care about the thousands of ICE cars that burn to the ground every year.
You even see it with seatbelts, people who claim it's safer to be thrown from the car in an accident.
But eventually those complaints lose their novelty and people just accept it as normal.
So five years ago, he was sure we were five years away. Now he's sure we are five years away. How is this a "more measured prediction?" Because instead of completely generalized self-driving used by everyone, he's predicting... completely generalized self-driving used only by companies?
I predict that five years from now, Umson will have a more measured prediction that is still somehow five years away.