Tesla has stated publicly they believe that within three years their cars will be capable of full autonomy, and expect it to take another two years to receive regulatory approval. By fully autonomous they state you will be able to walk outside, have the car approach to meet you, it will open and close the door for you, you can fall asleep, and several hours later wake up at your destination.
Yes, Tesla is taking an incremental approach to releasing the feature sets that are required to have a fully autonomous vehicle, but no, the end product goals for Tesla and Google are not different in kind.
What certainly is different is the manufacturing approach the two companies are taking. Google is seemingly aiming to release a fully autonomous vehicle at version 1.0, meaning every system of the car, such as manufacturing process, sales, customer support, will be at version 1.0 at the same time. In contrast, when Tesla releases its version 1.0 of the fully autonomous driving feature set, they will already have very matured versions of the other components, such as their manufacturing process, battery and drive train technology, sales and marketing, customer support, etc.
Plus, the Google cars look like something one buys for their four year old niece or nephew.
"...the Google cars look like something one buys for their four year old niece or nephew. "
You have to stop thinking like some guy from Mad Men and more like somebody buying AWS instances.
Imagine that instead of buying a single car that you drive everywhere you instead reserve a car for your daily commute, you might get various options, eg:
$500/month - Tesla Sports car, 30min journey, unshared occupancy
$300/month - 2 seat Google-Car, 30min journey, unshared
$125/month - 8 seat van, 40min journey, shared, up to 1 vehicle change
Now if you are a go-getter gunning for VP you'll pick the sports car. But others might not see the extra expense as being worth it.
First, I'm pretty far away from "some guy from Mad Men". That's actually quite a reach into my psyche; it would be interesting to see how you came to such an incisive reading while knowing so little about me.
But as far as the business model, as someone that does not own a car but instead utilizes a mix of public transit, bicycle, ZipCar (sometimes renting the VW Golf, sometimes the Jeep, etc.), and Uber to get around, I still think the Google car looks like something one buys for their four year old niece or nephew, and have faith that a car could check all the boxes functionality wise and still be designed to look, nice, at least. Maybe even really nice.
> $125/month - 8 seat van, 40min journey, shared, up to 1 vehicle change
This is interesting -- it's basically a little autonomous bus, for essentially the same price as a regular one, that might pick up/drop off at a more convenient location, minus someone there to supervise. I kind of like having a bus driver.
But would you rather walk 3 blocks in the rain (or snow) to a bus, wait at the stop for it to come, then walk another 3 blocks to your office, or would you rather have the minibus pick you up at your door and drop you at your office? Point-to-point travel is where self-driving cars excel - even if you're sharing the vehicle with others so you don't have a direct trip.
>Tesla has stated publicly they believe that within three years their cars will be capable of full autonomy
Tesla has stated many lofty objectives, some of which they've met, some of which they haven't. These engineering problems are hard.
I agree with the article; what Tesla has introduced with "Auto Pilot" is an improvement on current technology, like adaptive cruise control and park assist. It's cool, but it isn't in the same sphere as what Google is doing with their autonomous cars.
Google is going for a cars as a service business, Tesla is going for a cars as a product business.
That means that Google is happy with cars that only work most of the time, and only goes to some routes, as long as they are fully authonomous and reliable when they work.
Tesla, by its side, is happy with cars that are not fully autonomous, as long as they work every time, and can go to any route.
Of course their products are different, and they'll not become alike if none of them change their business.
As far as I'm aware, Tesla is selling cars to people who want to drive cars. Real cars, with muscles.
Google isn't selling anything, and again, as far as I'm aware, it's not settled whether they ever will, for instance there's been a lot of speculation on a rent-a-ride model.
And the prototype they have is clearly aimed at people who don't want to drive cars, i.e. more or less exactly the opposite segment.
> will open and close the door for you, you can fall asleep, and several hours later wake up at your destination.
Not going to happen. Seatbelts and upright seats meant for crashes are not going away even if the robots do all the driving. And I have yet to see any attempt at a robocar capable of understanding and interacting with today's chaotic and haphazard parking rules.
Heaven help us should all passengers be required to sign user agreements and watch in-car safety vids, but that could be a reality. I've got a BMW that already asks me to "agree" a contract for the navigation system every time I start.
Wearing a seatbelt and sitting (mostly) upright is not orthogonal to sleeping - people sleep in barely reclining airline seats all the time, on long car trips my wife and I take turns dozing in the passenger seat, with seatbelt secure and seat back only slightly reclined.
Self driving cars don't have to park where there are haphazard parking rules, they can drive to a garage or lot that caters to self-driving cars.
"...the Google cars look like something one buys for their four year old niece or nephew. "
They really are quite ugly. I believe that will impact adoption rate.
Modern cars are styled to be aggressive because that's what sells. The Google car is almost certainly designed to look harmless, because they want people to be comfortable riding in it. My late grandmother, for instance, refused to buy a car that had an angry 'face' (i.e. the shape of the headlights).
>Doing many thousands times better will not be done by incremental improvement
This assertion isn't obvious to me. In my experience incremental updates are often exponential in their impact (especially if enough resources are put into a problem). Moore's Law is an excellent example of this: at any given time, researchers are working on a fixed number of solutions that will generally make a fixed % impact. This is why we can see a doubling in transistor density without a huge increase in the size of the industry.
In the case of reducing accidents, I could see a similar exponential pattern. The first incremental step maybe took the accident rate from 10% to 1% by eliminating 90% of the possible sources of accidents. In the second step, researchers will again shoot to eliminate 90% of the current causes of accidents, bringing the rate to 0.1%. This could repeat every couple years until the accident rate is sufficiently close to 0.
The Tesla "autopilot" is comparable to what BMW[1] and Mercedes[2] have been shipping for years.
Self-driving is much harder. The first-order problems of driving on a empty road were solved by the DARPA Grand Challenge, ten years ago. The second-order problems involve dealing with other road users. That's hard, and that's what Google is working on, with considerable success. So is the CMU/Cadillac consortium, which has demonstrated their self driving car to politicians in Washington traffic.[3] Nobody seems to pay much attention to that effort, although they may be closer to a production product than anyone else. (Or not; Uber hired some of the people involved away from CMU.)
Self-driving cars need and have a lot more sensors than semi-auto cars. There's a lot more sensing to the sides and rear, and more forward sensing than just being able to detect the next obstacle ahead. Vision processing is far more elaborate. Google's vision system explicitly recognizes humans and bicycles.
Google's little 25MPH driverless car is a way for them to enter the market. At 25MPH, slamming on the brakes is a good solution to situations the system can't handle. Those things are going to be all over senior communities in a few years. Google already has higher-performance cars on the road; they can be seen all over Mountain View most days.
Reminds me that we didn't hear about Google SDV for a while (the last news was the 'View from the seat ..'). I expected a little increase in progress reports.
I just youtubed 25mph frontal crashes and it doesn't look pretty...
I'm surprised the article didn't mention Mercedes' history with drive assist systems especially. This is very old tech for them. For instance, I recall seeing a Mercedes S class with dynamic cruise control in 2002.
Because Mercedes isn't making absurd claims about autopilot systems and because "everyone knows" that innovation only comes from Silicon Valley. OK, that was a bit snarky but assistive driving systems, however useful and safety-enhancing, don't get people all wide-eyed and excited the way promises of autonomy in "just five years" does.
I agree that Teslas system is nothing special and more of a marketing win. But Google desperately needs to ship something that isn't a cool video or a swell PDF if they want to be relevant in this space.
Based both on what I've read and discussions with people who have some personal familiarity with the space, the issue is that, to date, Google is apparently not interested in anything incremental. (Which is also the reason that they haven't been able to work with the auto makers.) However,that means that Google won't have anything commercial for a long time--like decades--if their criterion remains robo-taxi levels of autonomy.
Almost-automatic driving is dangerous. There's a "deadly valley" which begins at the point where the driver starts to no longer pay attention. It ends when the automation is good enough that the driver doesn't need to pay attention. Between those two limits lies trouble.
The major vehicle manufacturers which have shipped driver assistant systems have put in controls that insist the driver pay attention. Mercedes, BMW, Tesla, and Ford check for hands-on-wheel.[1] They're desperately trying to stay out of the deadly valley.
Here's the problem with that.[2]
Volvo, and maybe BMW, have a second problem - too many modes and too many options. There are lots of semi-auto options available, and they may or may not be installed on any given vehicle. Here's what happened with someone who ordered self-parking without buying the "pedestrian detection" option.[3] Bad idea selling that option combination. It's been a learning experience for Volvo.
What's the problem with [2]? The driver knows he's bypassing safety interlocks. There many dumb things you can intentionally do in a car; adding one more is not necessarily a significant problem.
> A full robocar product is only workable if you would need to correct it in decades or even lifetimes of driving.
I had a conversation about this with friends in Germany a few months back.
In most societies, a mistake that causes suffering to another individual is usually 'blamed' on the person causing the suffering. In many cases where causality is obvious, this assignment of blame is fairly straightforward. Example: Bob fell asleep, which caused him to lose control of his car, which hit the bus, which killed a child. Bob is now culpable for the child's family suffering. Bob remains one of many others who share culpability at this point, assuming others are also falling asleep at the wheel. FWIW, 103M people fell asleep at the wheel last year in the US, so Bob will likely have company.
Now put an autonomous piece of software written by company X into Bob's car. Bob engages the autopilot, falls asleep, the autopilot software experiences an error, the software fails to alert Bob, the software loses control of the car, which hits a bus, which kills a child. Who is culpable for the family's suffering now? The software? Company X?
The only way for company X to both a) allow Bob to fall asleep and b) bear the culpability for a family's suffering is to get the software to the point it only makes mistakes in a timeframe that is, at a minimum, several orders of magnitude greater than Bob making the same mistake.
The logic goes that, once a company's software kills a child, it's going to be pretty hard to keep the public from reacting negatively, even though overall suffering will decrease. The only option company X is to require Bob to accept he is "driving" the car and bear the culpability of any suffering the car's software may cause, or alternately, be ready to pay a substantial settlement that offsets suffering.
It's an interesting question in that properly maintained and properly driven automobiles do have accidents that are clearly no one's fault--skidding on a patch of ice, deer running into the road, etc. Perhaps a more skilled or more cautious driver could avoid such an accident--or not. I'm hard-pressed to think of many other examples where a product used as intended will nonetheless cause harm to the user or others with some finite probability--but aren't considered the fault of the manufacturer. Pharmaceuticals and other types of medical equipment probably come closest. (Though drug companies certainly face lawsuits for side effects all the time.)
Chainsaws and handguns are probably other examples.
Fast food is another. (Alcohol, tobacco?)
Side note: I would argue that your patch of ice example is not nearly as good as the deer one. Skidding on a patch of ice and crashing is, IMO, simply driving too fast for conditions.
Fast food etc. though is more "stuff that's bad for you taken to excess" as opposed to something that can randomly get in an accident though.
I agree the deer is the better example. You can have patches of unexpected black ice though. I've skidded a few times though never had an accident.
There's certainly gear one uses in activities that have some inherent danger like your chainsaw example. I guesxs things like skis and helmets could be another. The difference with an AI though is that it's the machinery itself that is making the decisions.
Liability will not depend on current laws; it will be determined in legislatures after intense lobbying by manufacturers. They will claim that without liability protection, the government will stifle innovation and economic growth.
Perhaps clauses banning class action suits and requiring arbitration will help them:
> get the software to the point it only makes mistakes in a timeframe that is, at a minimum, several orders of magnitude greater than Bob making the same mistake
That seems to be more or less the case with the Google cars. 300,000 miles, with no accidents caused by the cars. Of course, so far that's limited to more-or-less good weather so far.
Unless you think Bob was causing accidents every thousand miles, I don't see how Google cars are "several orders of magnitude" better. Leaving aside the good-weather/known-roads aspect, of course.
The average accident rate in the US is something like one per 250,000 miles driven. (The average driver drives 15,000 miles per year, and goes about 18 years between accidents.)
So indeed, we can't draw much conclusion about the safety of Google's car relative to the average driver yet, except to say that it's not catastrophically worse.
I've always wondered how autonomous cars would handle the first and last quarter mile of the journey. I'm talking specifically like portions of the trip like the driveway, getting out of the parking ramp, or navigating small alleyways where the car could be parked (where GPS could be weak in the city). Things like even knowing which entrance to go to. Will fully autonomous will ever be able to take us from A to B 100%? Will humans always take over the last tiny bit where the maps aren't detailed and to park? Humans love to drive around the lot to park at exactly the "perfect" spot. Cars can parallel park now, but how will cars decide where to park exactly? Will we ever be able to have the car take us through the drive-thru?
I personally think autopilot-like auto-cruise just on the highway and more established local roads would be good enough. The convenience afforded by having the robot take us from A to B parked to parked may not be worth the insane price it must have on its tag to get there.
The car can drop you off at your destination and then go park wherever it's convenient. It can take as long as it wants since you're no longer waiting on it, so it doesn't have to worry about navigating "unparkable" areas. It could go to a networked garage where the cars stack side by side without space even for the doors to open. If your visit is short it could just drive around the block a few times. Alternatively, it could go pick up another passenger.
> I'm talking specifically like portions of the trip like the driveway, getting out of the parking ramp, or navigating small alleyways where the car could be parked (where GPS could be weak in the city)
Relatively simple from a conceptual point of view: a human drives it the first time, and later times use sensors to apply basic rules and the original 'instructions' learned from driving it manually the first time.
Extend this to a network of cars sharing route information, and only a small sampling of the population ever needs to drive any given route manually.
If Google's 25mph car is able to slam on the brakes and avoid an incident without swerving or taking other action then it could be argued that once we move to a fully autonomous society, 25mph (example; may not be accurate but for the sake of debate, 25mph is what I'll stick with for this scenario) may end up being the max speed for safety reasons.
This isn't to say that now trips will take longer and that will adversely impact our lives because I think what will end up happening is we will rearrange our lives so that we use these longer driving trips to sleep or work, converse with friends, do homework on the way to class, etc. and thus the time it takes to get from point A to point B becomes moot as we are now able to be orders of magnitude more productive in our vehicles.
Granted, this will not only reduce accidents as now the vehicles can communicate with each other and will instead know what the intentions of the other car are and adjust accordingly instead of trying to anticipate what the other car is going to do, but it will also reduce or eliminate speeding tickets and DUIs. Due to speeding tickets and DUIs being a large source of revenue for municipalities, I'd expect this to evolve as well, unfortunately.
The thing the author is missing is that Google can't incrementally improve, since they're not in the car business. Tesla, on the other hand, has the option of either incrementally introducing autonomy to their cars or taking the Google approach of shipping a 1.0 in a big release years down the road. That they've clearly chosen the former is telling.
The author pretends that both companies have a choice and have chosen different strategies, but it's clearly not the case. Unless Google was planning on building a traditional car business first (a fairly ridiculous proposition), or partnering and integrating with the supply chain of a major manufacturer (a stretch, if just to introduce fancy cruise control), they were never going to be able to iterate towards a robocar.
Guess what... Apple didn't sell the first smartphone either.
Someone takes a small step into car driving automation, tries to create some buzz around it, then I've got to read about how it's not a big deal. The nuances between autonomous and auto-pilot need to be discussed. We need a taxonomy.
I guess writing these sorts of articles is a million times easier than adding any autonomous features to any vehicle.
Forward progress is extremely important. It comes technically and socially. Let's hope everyone demands a car with "that stuff" they have in a Tesla. We'll get a little arms race that'll pay for further development, lives will be save (in total), and we'll asymptotically approach the vision.
"This is not a difference of degree, it is a difference of kind. It is why there is probably not an evolutionary path from the cruise/autopilot systems based on existing ADAS technologies to a real robocar."
I wouldn't take it as gospel, as it's hard to know for sure, and some companies are betting a lot on there existing that evolutionary path. Their argument might be: why not? Just because there are too many orders of magnitude of improvement? It's not that hard to think of examples where technology advances many orders of magnitude in small compounding steps. The author also mentions "new sensors" as if one can't add and remove new sensors in an evolutionary fashion as they are needed.
The author didn't justify that statement at all. If we have gone from an error once per minute (cruise control) to once per 30 minutes, and think that level of improvement can be repeated twice more, we will be at one error every 450 hours. A third time will put us at one error every 13500 hours.
Is continuous, gradual improvement the best way to fully autonomous cars? I don't know. But the author's argument is simply that we aren't there yet.
The trouble is that if they're not continually involved in the driving process, drivers can't actually concentrate well enough to be able to step in quickly when something happens that the automated systems can't handle. If I recall correctly, there have been studies on this and it takes tens of seconds for drivers to be able to respond to an unexpected situation sensibly if they haven't been actively driving. Mostly-automated cars that rely on drivers to step in when something goes wrong are probably not an option.
This interaction of computer and human decision-making is actually the subject of a fair bit of active research. e.g. at Duke http://hal.pratt.duke.edu/
To your basic point--yep, at some point you need to stop automating unless you're willing to hand over control entirely.
It's learning. That is an interesting approach. I wonder how far they get by that.
I guess Google's car will also collect data and help Google to improve the performance. My impression so far was that it's mostly engineered work however, and not so much learned (in a machine learning way).
Another way to improve the accident rate is the other side of the robocar argument wherein we, as humans, do a better job of driving and of watching/educating our kids.
I understand there are rules of the road and rights-of-way but a right-of-way for a pedestrian in a crosswalk with the walking signal is not going to stop a bus from running the light and killing the pedestrian.
Not that I'm blaming the pedestrian but it surely doesn't hurt to think defensively, look both ways and judge if that bus is going to be able to stop and act accordingly.
I'm always wondering how Google's car will certify when new releases are made. Consider that an autonomous car will need X hours/miles of driving before it will be certified. Now, if Google updates 1 line of code, the whole certification process has to start all over.
I would think the certification might not exactly look like that. At least, they aren't going to release every revision if the certification is like that.
I sort of think they are collecting the super detailed sensor data so that they can play it back into new versions of their software and see if it notices things earlier and makes better choices and such. A mass market version needs to be able to function on environmental data (like a tree across the road), so I don't think they are building a perfect map to have as a crutch.
Yes, I thought about replaying the information too. But the problem is that every different reaction by the car will lead to new input from the environment. And you can't store what you didn't anticipate :)
A hundred thousand times better is only seventeen doublings.
Which approach has the fastest exponential growth curve? The one with thousands of cars on the road, learning from each other, or the one with a few but more capable cars? We'll see. Just remember to think exponential not linear.
The point everyone is missing is the reaction time of a driver compared to a computer is so different. A computer have millions of cpu cycles to estimate the best decision during the time a human haven't even understood there will be an accident.
You're comparing # of cpu cycles to time taken to "understand" something. If the CPU can make a wrong decision REALLY REALLY quickly it's still gonna be wrong, whereas the human will have a continuous gradient of moving towards decision making.
You people crack me up. Self driving car was never anything other than an elaborate PR ploy for Google, the company which derives the vast majority of its income from its advertising business. Who would you rather work for: a company that is "building a self driving car" or a company that tracks the living shit out of everything you do on the web and ruins the web with mostly irrelevant ads? That's what I thought.
And they want these PR gravy trains running as long as humanly possible, so launching a real product isn't even a goal.
Yes, Tesla is taking an incremental approach to releasing the feature sets that are required to have a fully autonomous vehicle, but no, the end product goals for Tesla and Google are not different in kind.
What certainly is different is the manufacturing approach the two companies are taking. Google is seemingly aiming to release a fully autonomous vehicle at version 1.0, meaning every system of the car, such as manufacturing process, sales, customer support, will be at version 1.0 at the same time. In contrast, when Tesla releases its version 1.0 of the fully autonomous driving feature set, they will already have very matured versions of the other components, such as their manufacturing process, battery and drive train technology, sales and marketing, customer support, etc.
Plus, the Google cars look like something one buys for their four year old niece or nephew.