Thank you. This is a pet peeve of mine. Self driving cars do not exist, at all, yet, nor does anything that's really all that close but if you point that out on HN you'll be met with breathless disagreement.
Even more relevant, there's literally nothing in the historical record to suggest that the real breakthroughs when they come will come from a company like Uber.
Edit: Figures from 2015 indicate human interventions to prevent an accident were about once per 74,000 miles or maybe slightly worse than humans but not that much. http://www.roboticstrends.com/article/google_self_driving_ca...
The devil is truly in the details. You're covering 99.5% of all driving situations with 99.9% reliability? Well, then the remaining bit is going to be pretty devilishly hard! What do you do about cars that are stuck stationary on the the side of a bit of shoulder-less road? Here's what a human being does: They "sense" that something is up on the road ahead because the traffic patterns are different. They observe what the drivers ahead are doing, then start exercising caution concerning what's going on at the side of the road, maybe even to the point of hugging the other side of the lane out of caution.
Are "driverless" cars doing this now? No. Are you going to have to develop human-equivalent AI to do that? No. But is taking things like this into account going to take a lot? Damn straight!
http://www.theatlantic.com/technology/archive/2014/05/all-th...
The key to Google's success has been that these cars aren't forced to process an entire scene from scratch. Instead, their teams travel and map each road that the car will travel. And these are not any old maps. They are not even the rich, road-logic-filled maps of consumer-grade Google Maps.
They're probably best thought of as ultra-precise digitizations of the physical world, all the way down to tiny details like the position and height of every single curb. A normal digital map would show a road intersection; these maps would have a precision measured in inches.
Why not? The team at Uber that is working on this is largely comprised of researchers from Carnegie Mellon and former members of Google's driverless car team. Historical record suggests otherwise...
I mean they are competing against Apple and BMW, among many others, presumably, who have actually made a car, computer, or transportation device that worked.
That notwithstanding, the burden of proof is in the other direction.
Remember, a company is merely comprised of the people within it. If Uber has people that have joined from the kinds of companies that you mentioned, it becomes more and more indistinguishable from them.
Not really. Software companies are like that. They are the exception to the general rule that companies are composed of people plus things like factories, equipment, patents, trade secrets, inventory, real estate, global supply chains, and dealer networks for service and support.
Making cars, even regular ones, is hard and Uber has never really done anything like it.
No they haven't.
I know this because there aren't any cars in the world that can self-drive.
But if you define what Tesla already has in production as a self-driving car, then you would be wrong.
Brace yourselves. AI Winter Is Coming.
Indeed. The unpredictable behavior of humans occupying the same environment is key. Remove this element and things get simpler (though not simple. And v2v comms might be required.)
I suspect highways are the road infrastructure most akin to a closed track. Which is why the Otto acquisition is so interesting.
https://en.wikipedia.org/wiki/Autonomous_car#Classification
Level 4 still a LONG time coming. That was underscored for me by this interview:
http://thisweekinstartups.com/george-hotz-commaai/
George is kind of arrogant and brash... but even he says "I don't know anything about level 4". Everyone thinks level 3 and 4 are the same thing, but now that I've heard him dismiss level 4, that actually gives me more confidence in his startup.
Those are two completely different things!!!
Google has been working on self-driving cars since 2009, and the DARPA challenges were around 2005-2007. We're not even at level 3. Level 4 might not come for decades after that.
Tesla has "good enough" L3 already on the road, and I highly doubt Tesla's Autopilot is anywhere near as sophisticated as what Google is working on in their labs.
You can get on a bus that does that now in Helsinki. Admittedly they are slow and on set routes but even so they're there.
http://money.cnn.com/2016/08/18/technology/self-driving-bus-...
Problem is, that makes them rigid and unable to handle anything unplanned for.
I can take a bus or train anywhere at any time, knowing that if I'm, for example, south of Washington at any time of day or downtown after 2am, I can call a car and still get home. In the 00s, the only way out of those zones was to bring your own car or be prepared for a "Falling Down" style 10 mile walk.
Yellow cabs (then and still) don't enter a massive buffer zone around regions where they might get a black guy as a fare even if you called, but Uber serves those areas reliably, 24/7.
Wherever it's deployed you have to conisder the traffic and congestion repercussions if everyone uses it, not just rich people.
Transit has to move poor and middle class people too.
But we did just spend twenty billion dollars on a bullet train from Bakersfield to Merced, so there's that, I guess.
I think it's a fair bet that autonomous cars will come before Geary BRT is completed, and it's a certainty that they'll come before Geary gets a subway. And while Geary is one of the worst, there are a half dozen similar bus corridors in the city. Also, most of the rail is above ground and similarly slow and crowded.
While I generally think rail and autonomous cars can be symbiotic, autonomous cars are likely to kill buses. Middle-income people ride the bus only out of necessity. High-income people never ride the bus, though they will ride a decent subway system. Once autonomous cars and autonomous shuttles come out, middle-income riders will flock to them, the only people riding busses will be the poor and working class, and the system will collapse, sooner in most other places, but it'll also happen in San Francisco.
Because San Francisco doesn't have a real subway network, SF MTA is in for some rough times in the not too distant future. I wouldn't be surprised if San Francisco bans or heavily taxes private autonomous systems. If they were smart they'd be digging tunnels as fast as possible. Some cities, like Madrid, seemed to have learned to do it cheaply, and the savings aren't just because of geology.
I've accepted the fact that public transit systems here will never be great. It will only be ever be barely good enough. There are a ton of factors affecting this, but at this point my faith is with the private sector.
I'm in Tampa most of my time, and it will be impossible to service that much sprawl with light rail or bus service. And if you don't like sprawl, you're going to need to change public policy to deal with dense real estate costs.
There is no free urban planning lunch.
Dense urban areas are better suited to mass transit.
Uber fills a niche of point to point transport that's current overutilized because of VC subsidization.
Even if you get rid of the driver, you can't expect everyone in an urban core to do point to point transit, especially with the inefficiencies of pick up and drop off.
i agree that at some point low-density sprawl wouldn't make centralised public transport economic (even supposing global regulation existed to force the cost of pollution to be accounted for in market prices)
at this point in our global human predicament it'd be a bit of a shame not to realise that some ideas that were okay in the past aren't very good ideas any more.
In my opinion, self driving cars are closer than the naysayers think. The cars don't have to be able to handle every situation initially. Driving from my house to the local shopping centre is only 4 turns, this could be a preprogrammed route to a designated autonomous-vehicle drop off point.
I'm going to say we'll see a lot more of this in the next five years.
That's similar to what I've wondered about as a good intermediate step between what we have now and fully autonomous self-driving cars that can go from any given place to any given destination given the address, even if they have never been there before.
The idea is that most of us have various trips that we make frequently, such as to/from work, to/from a grocery store, to/from kid's school, etc., and maybe we could simplify the early consumer self-driving systems by limiting them to such trips. Here's how I imagine it would work.
When you decide you want to add a route to your car's self-drivable routes, such as your commute to work, you put the car in an observer mode and drive the route manually. The car tries to identify traffic signs, traffic lights, lane markers, and other such things that it needs to understand. The car also records video of the drive.
When you tell the car you've reached the destination, it uploads the video of the drive to the vendor that provided your self-driving system. It also uploads what traffic signs, lights, etc., it has identified, and a log of your control inputs.
The vendor has a human watch the video of your drive, review the things that the car identified, and identify things that the car missed that are relevant. The human accesses whether or not the car can handle this route as is, or can handle it if some special cases are added, or cannot handle it. In the first two cases, the vendor notifies the car that it is now allowed to drive the route (and downloads any extra information they identified that the car needs), and the car can then drive the route in self-driving mode.
Whenever the car drives the route in self-driving mode, it compares what it sees and identifies (traffic signs, lane markings, etc.) with those that were used when it was shown the route. If there are more than trivial changes, it asks the passenger to take over driving, and uploads a report to the vendor for review.
Perhaps it also uploads video of the route every time it is driven, and the vendor runs software that automatically compares each to prior videos of that route, and learns what features normally vary over time, and which are more constant. The results can be sent back to the car to adjust its thresholds on a feature by feature basis for when it considers something changed enough to require human intervention.
Who's going to preprogram it? And then the billions other potential preprogrammed routes.
You have that scope reversed. That's why multiple organizations have autonomous cars, which have been real world tested, to various success, and are being allowed to roll out (the whole derail of "they don't exist yet" is naysayer noise) without nine nines reliability. While the moon trip was only done ONCE across all of humanity's history and there wasn't a test beforehand. THAT was impressive, to the point of incredulity.
> But since the technology has not been perfected, the cars will come with human backup drivers to handle any unexpected situations.
Getting those last 0.01% of situations right will take a long time.
I don't think the software is the gating item for full autonomy. The production cycle of the hardware (the actual cars) is what we'll be waiting on most over the next few years.
In a city, not only do you have better/more frequently updated map data, but you also have a large corpus of cars that can share info with each other, which will be a big aspect of driverless cabs. Also the speeds don't get as high, so safety is a bit easier. Also flooding/snowstorms/other things computers can't yet handle well are more of a problem on the open road.
Your assertion that moving freight must come first, to me, has no apparent logic behind it.
If you think about it, taxi cab drivers barely need any qualifications to drive a cab. Truckers need more, almost as if trucking is a more demanding type of driving. By that alone, I feel like it's fairly apparent which would be easier to replicate on silicon.
Sure it does. It is the easier problem, and it makes sense to assume the easier problem will be solved before the hard one.
Self driving cars that move about long stretches of interstate highways with little to no interactions with intersections, pedestrians or complicated traffic patterns is a far easier problem than driving in densely populated urban areas.
If you think about it, taxi cab drivers barely need any qualifications to drive a cab. Truckers need more, almost as if trucking is a more demanding type of driving. By that alone, I feel like it's fairly apparent which would be easier to replicate on silicon.
Trucking is a far more demanding job than driving a cab, but that is only true for humans. What makes trucking demanding is long, rigorous driving schedules that require high degree of driving stamina. As it turns out computers have very high stamina and are very good at doing long rigorous activities without suffering any fatigue. This is as opposed to driving a cab where the difficulty of the job is not so much in the rigor of the job but more in what humans do better than computers. Which is mainly dealing with other people, understanding patterns and learning from past experiences.
Additionally, they need to be able to drive at night in areas without much lighting, they need to drive for hours and hours on end, instead of just 5-45minutes, etc. Finally, they are 40+ feet long (instead of being around 15 feet long) and have much less maneuverability and a much longer stopping distance/time.
Sounds like a harder problem to me.
Freight consists of both long and short haul, with different needs and approaches. Your point applies to the "last mile" delivery phase, which may or may not be a significant part of the operation.
In fact, many big stores like Walmart, Costco, Aldi, and the like are placed pretty close to an interstate exit. I would guess that this allows significant reduction in costs, as the operators don't need to switch from their optimized long haul vehicles (big trailers, etc) to more maneuvarable, less efficient (per unit weight of goods) trucks for the last mile stage.
There is plenty of freight that is carried on highways over long distances with low traffic volumes, in a similar role to rail freight.
> Additionally, they need to be able to drive at night in areas without much lighting,
During winter at higher latitudes a lot of driving is at night. As for lighting, plenty of city streets are dark and have poor lighting at night.
> they need to drive for hours and hours on end, instead of just 5-45minutes, etc.
So? Fatigue is something that affects human drivers. There should be no serious differences for an autonomous vehicle driving 10 30 minute trips and one 5 hour one, modulo start and stop. In fact, if anything, longer duration trips tend to involve a greater fraction of highway usage that does not require significant brake usage, turning, etc.
> Finally, they are 40+ feet long (instead of being around 15 feet long) and have much less maneuverability and a much longer stopping distance/time.
I agree with this.
You could create a system, potentially using truck stops or in some way your own infrastructure, which is a docking point for autonomous vehicles. At these locations, human drivers pick up and drop off the otherwise autonomous vehicles, allowing humans to take the burden of the difficult portion of driving, while computers handle the boring part of driving. Of course, you would have to place similar stops every (80% of a tank of gas, measured in distance), as well for refueling. Naturally, this idea has a dozen holes at the moment and would need to be improved to be viable.
Driving in weather is harder, but that is still a problem for city cars, I'm not sure why you think it is more of a problem on the open road.
The map data might be more frequently updated, but it is also more complicated. Stop and go traffic (with lane changes!), turns, traffic lights, construction, busses, etc. all add to the complexity of city driving.
The first thing for freight isn't going to be completely autonomous, it is going to be a slightly smarter tesla style autopilot, i.e. a driver who is alerted by the computer when he needs to take over (weather, pulling into towns, etc).
I honestly have the opposite opinion that you do: freight (ignoring the last mile) is a much easier problem than inner city driving.
Furthermore, those trucks are expensive. Tacking on an extra $10,000 to a family sedan for sensors and electronics is huge. It's not so extreme when you're paying $100,000 for a base truck.
And finally, those fleet vehicles have regular inspections and maintenance. Everything works great when it's all shiny and new. It'll be interesting to see what happens to autopilots on cars that don't even get regular oil changes.
Well, part of those rules and regulations do things like specifically limit how much they can drive per day (11 hours is the limit, I believe). There's also super high turn over in the trucking industry and they've had driver shortages fairly regularly. A lot of these rules are about protecting the human driver from themselves (fatigue, misbehavior etc). When I think about the problem it seems that regular, predictable freight movement from port to warehouse would be a logical place to start because a) it's easier and b) the labor market for drivers is so severely distorted right now.
Edit: It's 11 hours but I'm remembering there's other restrictions too, like limits on consecutive blocks of 11-hour driving without rest. I'll look it up later.
But as we get closer and closer to that inevitability, the value of Lyft and other non-autonomous ride-sharing companies will just get lower and lower.
I actually think because the rewards are so great, and the technology is so difficult, that infrastructure will change before the tech. There are things you can do to drastically reduce the complexity, like install smart traffic lights and have cars talk to each other, in addition to special road markings designed for self-driving cars.
CMU has been working on self-driving cars since 1984, people seem to think nobody was doing anything before Google (and they also think that Google's cars are far more capable than they are). Yes, money makes a difference, but sometimes reality just disagrees with profit. And if there's something startups are good at, it's hyping themselves up to jack up valuations.
And after all, we only really need a statistical improvement over human drivers for it to be worth it, given that the software in these cars can be updated over time.
I guess it would be nice though, if we had a really tough standardized test course, involving various common situations with pedestrians, cyclists, etc.
On the other hand, we test human drivers on real roads too.
Currently, bus/overground/tubes/whatever have only few lines and they can cover a lot of users.
Tomorrow, self driving cars could cover only the few main roads of a city. That would be enough to be useful for 80% of people in the city. That ignores the hard 90% roads which are harder to do.
We'd still have to walk down the block to reach the uber and that's annoying, but that's a huge progress over the current public transportation systems (whom coverage vary hugely by city and country, to the point there might not be any).
This seems wrong. Wouldn't taxis be first, since cargo still needs to be hauled off and loaded by someone/thing? Many of the those drivers double as laborers.