What's the difference? Apart from the legal requirement to have a diver ready to take over in test vehicles (which necessarily makes it Level 2), the fundamental difference is that you'd have to show a lot more than one demo to establish that you've achieved Level 4. Level 4s are supposed to be able to operate without human intervention at all within prescribed domains (e.g. downtown cities). That doesn't mean operate one trip or one day or one month without a disengagement -- that's still Level 2.
I'm super impressed by the demo but Cruise will have to show more data to back up a Level 4 claim.
when we see the streets of San Fransisco conquered, then we know that self-driving is ready to come of age.
If you've ever been to those countries, the first thing you notice as a westerner is that the streets are chaos, traffic rules are barely obeyed, and aggressiveness is required to get anywhere.
Meanwhile, Beijing and Dehli have extreme pollution problems partially caused by cars.
If any place was in need of point to point autonomous ride sharing, it's Asia. The reduction in cars would reduce congestion from the outset. But in the future, protocol guided right of way negotiation could reduce congestion even further.
Ideally, the AI should be able to figure out how the dynamics of such a situation work and not create a jam or something.
I still haven't seen a driverless car handle a single-lane road (e.g. with parked cars on the side). Something like this:
https://www.google.co.uk/maps/@51.4606316,-2.5562767,3a,75y,...
A car forcing a merge right in front of it, an armored car taking a lane and a half where it had to wait to go around (there were numerous instances of similar lane blockages that it navigated around), a pedestrian jay walking in front of it, a bus picking up passengers, a bike cutting in front of it, and car pulling out in front of it in heavy traffic.
I had no idea they could handle that many scenarios already.
Notes:
* There are frequent steering twitches to the left. This may be associated with passing parked cars. There are similar twitches to the right when in the left lane of a one-way street.
* Crosswalk behavior when turning needs some work. The vehicle enters the intersection, then stops in the intersection before the crosswalk with people in it. This is a hard problem, because the system needs to recognize people waiting to cross but not yet in the roadway. When the light turns green, both the pedestrians going straight and the turning vehicle can enter the intersection, the pedestrians having right of way. The pedestrians now block the vehicle, and the vehicle blocks the bike lane.
* Left turns into multi-lane streets are too wide and into the wrong lane.
* On two occasions, the vehicle is stuck behind a doubly-parked vehicle engaged in loading. The options are to wait or to cross a double yellow line. There's a delay of several seconds, then forward movement. Suspect manual intervention.
That was nice and clean city driving in the video clips but nothing that distinguishes it from "Level 3" (human intervention may be required within ~15 seconds or so) or even "Level 2" (human intervention may be required within seconds, current state of the art).
The benefit of this is you notice where levels "poke through," e.g. Level 4 may work on a sunny day, but a small change in rain or road conditions could downgrade the system to Level 1. As time goes forward, the inner levels can be expected to radiate out.
EDIT: Never mind. The point of "Level 4" is it is competent in all reasonable operating domains.
This isn't the way it works. If a car says it can do "Level 4 on a sunny day" it means when you sit in the car and engage the autopilot (and it says it's safe to engage), then no human intervention will not be required during the course of the trip. If conditions change, the car will be parked without human assistance and wait for help (what happens next is outside of the scope). You could be sitting in the back seat, or your kids could be in the car alone on their way to school.
"Level 4 on a sunny day on some pre-certified highways" is ok. "Level 4 in San Francisco traffic" is also ok, and much harder. "Level 4 unless it starts raining and we'll deteriorate to Level 3" is Level 3, not 4.
This is the definition of autonomy levels from SAE, and they're pretty strictly defined.
Level 1: Driver must be ready to take control at any time. Automated system may include features such as Adaptive Cruise Control (ACC), Parking Assistance with automated steering, and Lane Keeping Assistance (LKA) Type II in any combination.
Level 2: The driver is obliged to detect objects and events and respond if the automated system fails to respond properly. The automated system executes accelerating, braking, and steering. The automated system can deactivate immediately upon takeover by the driver.
Level 3: Within known, limited environments (such as freeways), the driver can safely turn their attention away from driving tasks, but must still be prepared to take control when needed.
Level 4: The automated system can control the vehicle in all but a few environments such as severe weather. The driver must enable the automated system only when it is safe to do so. When enabled, driver attention is not required.
Level 5: Other than setting the destination and starting the system, no human intervention is required. The automatic system can drive to any location where it is legal to drive and make its own decision.
From the Society of Automotive Engineers
I'm also curious how it would react to going up Mason and California, where there's a traffic light at the top of a steep hill. Last time I had to physically pull myself up via the steering wheel to see anything, and as a seasoned driver I was a bit worried.
Add to this complexity the weather conditions. Suppose the sun is shining straight at you and you need to squint and shade your eyes just to make out what the light is -- this happens to me frequently -- can the camera see the traffic light and distinguish its color clearly under such conditions?
What about when it's raining, misting or drizzling, snowing heavily, etc. and the traffic lights are these fragmented outlines that you, the human, can heuristically distinguish but a machine might not?
One last thought: suppose it's right turn on red and first car in line is a self-driving vehicle. Can it really look left and safely determine there's enough time to beat the cross traffic? If it's highly conservative and just waits until green, there could be ten irate motorists behind it and guaranteed to honk and curse.
It's exciting technology but there are some very difficult problems to solve. I worry that if these machines can't demonstrate 110% of a human's ability to drive, they simply won't be implemented in many places except some very well defined rigid routes that are free of problematical challenges and variations.
I couldn't find any real information on this though. I think I picked it up from Reddit comments. Some info does suggest it's legally complicated: http://www.claimsjournal.com/news/national/2014/03/06/244965...
It's not an impossible problem to solve, but I have not heard of any efforts to create a car-to-car or car-to-road communications protocol and infrastructure that would be cross-manufacturer and/or internationally approved.
In my opinion, before such a mechanism exists, autonomous cars will only work as long as the majority of cars are driven by humans. I will change my opinion when I see a demonstration of a city (or city-like test site) traffic where all or the majority of cars are autonomous.
Nor do they have any presence I'm aware of on Github. This is in contrast to BMW, for example, who have made a number of contributions: https://github.com/bmwcarit
Anyway, just curious to what extent Cruise used (or still uses) ROS and open source software in their stack.
You might not be able to turn an aircraft carrier on a dime but when you do you've got an aircraft carrier.
GM (and the other automotive manufacturers for that matter) decided they wanted in on self driving and electric vehicles, had a few meetings, wrote a few checks and a few years later look at the result. Are Google and Tesla going to reply with similar videos?
I'm not sure what this means.
(Southern Michigan gets quite a lot of snow. Prevailing winds are westerly so lots of lake effect. Grand Rapids gets almost 2x the snow that Detroit gets. Cleveland also gets more snow than Detroit, lake effect from Erie)
Every time ... cars pass it, and dive into its lane (a typical reaction to slow moving vehicles).
Still, nice accomplishment.
Thought: Once we have 99% self driving cars it will be quite easy to convert a portion of roads to pedestrian only at times when traffic is light: bollards go up, lighting changes, cars informed to reroute.
And so you end up with posts like this trying to analyse a video frame by frame to assess the reality of the technology, and yet everyone including the author tries to guess where's the catch ( is the green light really trustworthy ? Why is the video accelerated ? Etc..).
I'm reasonably certain we can stop right there. It's a heavily regulated field and these things can't be sold until they've been rigorously tested. I know SV isn't known for its coziness with regulatory agencies, but in the field of safety, they'll have to play ball or take on all the liability the lawsuits will throw at them.
Entering the bike lane appears to be legally-required [1] behavior for a driver turning right across a bike lane in California. Entering the intersection, however, i'm not sure about.
Illustrated here: https://www.sfbike.org/news/bike-lanes-and-right-turns/
[1] https://leginfo.legislature.ca.gov/faces/codes_displaySectio...
Near my home there are several intersections with bike lanes, and I estimate that perhaps one driver in 10 enters the bike lane to make the right turn as they are required to.
In fact, I sometimes get surprised or even dirty looks from drivers making illegal right turns when I do move into the bike lane to make the turn. I guess they think I'm trying to get ahead of them and cut them off, but I'm not, I'm simply following the law and improving bicycle safety.
Worth mentioning, 30 seconds would feel like an eternity inside a car.
Google had another crash last month.[1] As usual, it was probably the other driver's fault. Other driver apparently botched a left turn in a two-lane left turn intersection at Rengsdorf and El Camino. Google drives in traffic on the SF peninsula every day, and we know exactly how many times they've crashed. It's reassuring seeing all those miles with only the occasional fender-bumper. It's not like Tesla slamming into something on a freeway at full speed. Three times.
Still waiting for the CA DMV to post the 2016 autonomous vehicle disconnect reports. Those cover December through November and are due Jan 1st, so DMV should have them up by now.
[1] https://www.dmv.ca.gov/portal/wcm/connect/3d358211-3f0c-430e...
Would it, when the car is driving for you and you're free to putter about with a phone or laptop or book?
I always figured that this would be one of the greatest benefits of autonomous vehicles - perhaps the car isn't as aggressive as the human it's driving it would be, or as quick to get "unstuck", but the human won't care too much because they're too busy on their phone. In most cases, what's a few extra minutes when you're no longer actually driving the car?
I would love to see how the self driving contraptions handle this city at rush hour.
Well, how do you distinguish / identify a traffic light?
For me it's a combination of knowing the area, knowing what a traffic light looks like, and observing the behaviour of traffic around (mostly in front) of me (if the intersection isn't visible).
For an autonomous vehicle, they'd use the same methods plus they'd have the non-trivial added benefit of colleague robots feeding them updated information.
> Suppose the sun is shining straight at you and you need to squint ...
I'm sure the human eye + sunglasses + visor comes a poor second compared to CCD + mechanical + IR + (etc) can do.
> What about when it's raining, misting or drizzling, snowing heavily, etc. and the traffic lights are these fragmented outlines that you, the human, can heuristically distinguish but a machine might not?
Interesting use of the word heuristic there. If you can determine the heuristics you are using, then an autonomous vehicle can use the same.
> One last thought: suppose it's right turn on red and first car in line is a self-driving vehicle. Can it really look left and safely determine there's enough time to beat the cross traffic? If it's highly conservative and just waits until green, there could be ten irate motorists behind it and guaranteed to honk and curse.
This, and variations, frequently come up in discussions on AV. There's an implicit expectation that no one building these things has considered this problem (this is clearly false). There's two explicit expectations that once a tipping point of AV's are out there, a) there'll be a regulatory push to massively accelerate the adoption close to 100%, and b) inter-vehicle communication means this problem won't arise. In the short term these problems may occur, but to address your question, yes, I'd suggest a computer would be better able to predict if there's enough time to safely turn than most humans.
> I worry that if these machines can't demonstrate 110% of a human's ability ...
Which human would you pick?
Anyway, I wouldn't worry - none of these problems are intractable.
red, yellow, green, flashing red, flashing yellow,
no-turn-left red arrow, no-turn-right red arrow
My concern is more about the regulatory side -- before any of this stuff gets widely deployed, various authorities are going to have to come up with some standards, and for the most part nation states are spectacularly poor at doing this sensibly in isolation, and horrendously woeful at doing it in cooperation. I suspect we'll end up with cars & infrastructure that are not 100% interoperable across borders.
Better signage and more consistent (state to state) rules that make common sense would go a long way.
Keeping in mind what is common sense to a biker is many times counter-intuitive to a driver. I hate the unprotected bike lanes on the right - as a driver you simply are not expecting people to be passing you on the right, on what effectively "feels" like a shoulder. It amazes me there are not more deaths due to this, to be honest.
After I spent some time in the Netherlands and Belgium I realized this was not just me. Those systems actually work and are designed with "road sharing" in mind. Everyone gets where they need to go, and even a dumb foreign tourist who knew nothing of local traffic laws could operate a vehicle safely amongst cyclists.
Intersections are the most frequent place for cyclist/driver collisions, and "right hooks" (where a driver fails to merge or yield and turns right into a cyclist trying to continue straight through an intersection) are the most common kind of intersection collision, so your instinct is spot on. Your perception that there aren't that many fatalities is probably partly because traffic deaths don't get reported on much, and partly because collisions involving turns tend to take place at relatively low speed, so fewer of these collisions are fatal than might be the case under other collision circumstances, as speed of the car is by far the strongest predictor of death in a cyclist/driver collision.
Having driven for a bit in California I'm now used to expecting people to be passing me on the right, on the left, between lanes - everywhere - thanks to a combination of legal and illegal lane splitting by motorcycles and mopeds.
But a broken traffic light is a situation that is unfamiliar to drivers, and at peak hours the crossing is guaranteed to be either difficult to navigate or to have suboptimal traffic patterns (otherwise there would be no traffic light in the first place). And of course pedestrians have to improvise (and Germany has a lot of pedestrians compared to the US)
For the record in East Lansing we get a lot more snow than Detroit but we're pikers compared to say Holland or Grand Haven.
Highway autonomy will be a nice feature but you still need a competent human driver to be present for end to end transportation. So I still need to order a car with a driver to get to the airport.
I assume that you're referring to trucking. It's possible that trucking companies will set up depots right off the highway and will reduce drivers as a result. But a lot of long distance container transport is already done by trains so it's not immediately obvious that eliminating drivers for only the highway portion of trucking is necessarily a big enough win to make it happen. (Especially if autonomous driving systems allow for the driver to sleep in his cab while underway.)
It's still domain dependent, right? If the rules say we're flattening precipitation and visibility within normal bounds, fine, but sometimes you have abnormal weather and badly-maintained roads. It is useful to compare hypothetical cars that autonomously navigate conditions no humans would dare.
EDIT: Never mind. The point of "Level 4" is it is competent in all reasonable operating domains.
No, it's not.
In Level 4, if there's "abnormal weather and badly-maintained roads", the car must be able to deal with the situation and enter a safe state. It can say "please sit down and wait for assistance" but it may not say, "take over the wheel".
In practice you'd probably disengage the autonomy and drive yourself, you take control and it does not give you control. If you can't sit on the back seat drinking beer or send the car to drive your kids to school (we're not considering legislation issues here), it's not Level 4.
This is an important distinction. "Level 3 in most conditions" is good enough to pass as "full autonomy" for most people, but Level 4 is a requirement for not absolving the humans on board for any legal responsibility.
To be fair, Tesla is not selling a fully autonomous vehicle, yet, people are treating it like one.
To provide proof of Level 4 autonomy with video (in certain conditions), it would need to show adverse and exceptional conditions, such as terrible weather, accidents ahead, all routes to destination blocked or any other situation short of a force majeure disaster and then provide a safe contingency for that.
In San Francisco, that would probably mean finding a parking lot where it is safe to wait for assistance from your Transport Service Provide(tm). In rural northern Europe where I'm from, it would mean parking on the side of the road and calling your wife/mom/friend to pick you up :)
Riding bitch through Hangpu to Pudong with a laptop in one arm while lane splitting with aggressive (and otherwise bad drivers) all around was a rather exciting experience.
Even with a skilled pilot, the adrenaline rush was similar to my first time tandem skydiving.
following lines from neal-stephenson's cryptonomicon come to mind:
"like Disneyland with the safety locks taken off"I'm going to get down voted for this, but it has to be said.
The whole "well it couldn't drive in X country" scenario is ridiculous. Driving in unruly, chaotic, unregulated environments is NOT a problem for self driving cars to solve.
That type of driving environment certainly IS a problem, but one that has already been solved by good governance, including laws, rules, and regulations for improving safety of traffic. Countries like the United States, among others, have developed a set of rules and regulations that make self driving cars possible within their domain. It's foolhardy to assume self driving cars can, or should, work outside of that domain.
That is of course, unless, you DO solve it under other constraints. Sure, you haven't solved X problem with Y solution, but you've still solved X problem. Self driving cars may not be the solution to ALL kinds of traffic problems, but it turns out, they don't have to be.
Too many people, myself included, that's all that matters.
They need to stop calling it "autopilot" until you can get in and fall asleep in the back seat.
In the medium term, I expect self driving cars will instead start to communicate with each other to find a better approximation of optional traffic flow (by finding the solution that is best for the negotiating vehicles). This has several competitive advantages over the simple algorithm: it enhances safety for participating vehicles by removing prediction errors, and if you can negotiate with the other car instead of guessing what it will do you can do much tighter manoeuvers with smaller seaters margins. And of course better traffic flow makes driving more pleasant, which is good for car sales. Given all those market forces I think it's very likely to happen.