The original article from the SF Chronicle has some more info. Specifically, it contains the following very strange paragraph
Sources said that Waymo does not plan on operating its cars remotely — even in difficult situations — but it will remotely monitor them during tests. If one of the cars encounters something it doesn’t understand, such as complicated road construction, the car will contact Waymo for help recognizing the situation. After human testers give it feedback, the car will then decide how to navigate the situation.
It seems like the round trip time for that type of operation would be pretty challenging for situations where you have, for example, six seconds to avoid crashing into a road barrier..
I’m not sure throwing control to a remote human in this situation would be much better given that they’d need to be able to immediately act correctly. I guess they’re confident about their abilities at speed. I would guess this is probably more to address “there’s an ambulance parked in the intersection, can I go around it or should I turn around and find another way” type questions
I don't think it's for that kind of situation. It's probably more for, like, a weird construction setup where the car needs to be told when it's okay to cross over the normal lines painted on the ground, or a parking garage where the car has to be told where it's okay to park (e.g. different sections may have different rules).
This is not video game style remote driving. The car is always in control, and its primitive object avoidance routines won't let it crash into a barrier. This is presumably to allow humans to help make higher-level decisions. For example, maybe the car asks something like "this road is blocked. I could cross the double yellow line to get around, or I could reverse and go a different way. Both of these options are suboptimal, can you help me choose?"
The most extreme type of assistance could maybe be a human drawing or modifying the vehicle's planned path (if you watch any of the Waymo example videos, you'll see the immediate path that the vehicle is planning to take). But even in this scenario, the car is still responsible for navigating that path and avoiding any obstacles.
> The car is always in control, and its primitive object avoidance routines won't let it crash into a barrier.The car is always in control, and its primitive object avoidance routines won't let it crash into a barrier.
Of course, how inconceivable that a self driving car could crash into a static barrier!
While I agree, yes it has been proven with Tesla's technology that it could happen, I believe that Waymo would not be so callous as to apply for this permit if it had not 100% proven that its technology would not do something as blatantly bad as that. I could be wrong, but I certainly feel through their various reports and showcases that they have by far and away the best grasp on fully autonomous vehicles. I have yet to read that in their millions of miles of fully autonomous driving there was a crash that was worse than cosmetic damage to either vehicle or caused injury [1]. They also seem ready to take responsibility for any of their mistakes and use them as a learning experience to further improve their technology.
> I believe that Waymo would not be so callous as to apply for this permit if it had not 100% proven that its technology would not do something as blatantly bad as that.
It it could prove this 100% to itself, it should be able to prove it to the general public and the authorities in a test environment..
What if for 8 hours a day your job was to avoid crashes in remote vehicles... every 6 seconds your monitor changes feed to a near accident and you have to react, your decision saves the vehicle and passengers or there is a crash.. next.. next..
If one of the cars encounters something it doesn’t understand, such as complicated road construction, the car will contact Waymo for help recognizing the situation
Hopefully well in advance of actually hitting anything, but therein lies the problem: what does it do if it's moving at speed and figures out that it doesn't know what to do? If it continues moving, it might hit something, and if it tries to stop suddenly, it might get hit and cause a multi-car chain collision. Imagine you're following a car on a highway, in a situation where no other cars are around, a situation in which a human driver would practically never stop, and suddenly it brakes to a stop very rapidly because of some perturbation in the road that confuses it, but a human would have no trouble with. Drivers are taught to look beyond the car in front and try to anticipate what it might do; and something so contrary to expectations is going to lead to problems.
In my opinion, the safe thing to do is probably not to "stop very suddenly", but to "pull over and stop in a safe manner, and have the blinkers on".
I think it would also be possible to mitigate this problem by not driving on the freeway at all for the first iterations. (Feasible if you only drive within Mountain View, and not e.g. MV <-> SF.)
What do you do if you're driving along and realize you don't know what to do? Do you slam on the brakes for an emergency stop? Probably not. A safer choice is to remove throttle input and consider gently applying the brakes while you figure out what's going on, and check to see if there's enough space behind you to brake harder or to pull off to the side of the road.
I would expect a fully autonomous vehicle to constantly be planning for emergency and graceful stops, and to slow down if it doesn't have a confident plan for both types of stopping and forward progress. It should also have plans for when critical equipment fails, such as sensors or actuators as well as normal car systems like brakes, tires, engine, steering.
People will die if GM/Cruise takes the safety driver out of the car. Anyone familiar with Cruise Automation's software will agree to this statement. Hopefully Waymo has done a better job.
Waymo does nearly 5500 miles per disengagement, GM seems to be at about 1200 or so as per the California submissions for 2017. Of course this does not tell us the quality of the miles - it is easier to achieve a lower rate of disengagement on the highway for example, but perhaps it serves as an indicator.
All others testing in California are 300 or fewer miles per disengagement.
>“We’ve staged people jumping out of canvas bags or porta-potties on the side of the road, skateboarders lying on their boards, and thrown stacks of paper in front of our sensors,” Waymo said in its report.
>For media day, they put on three of them. In one, a Honda convertible abruptly cut off one of the Pacificas just past an intersection, and the Pacifica braked to avoid it. In another, a car obscured by two parked vehicles suddenly backed out of a driveway. Again, the Pacifica came to a halt. The third was a faux moving-day scene, with a couch in the street and boxes tumbling into the lane, just as another car approached from the opposite direction. The Pacifica knew better than to swerve into the car. It braked until the coast was clear.
There should be an independent body to frame these tests (with enough random stuff thrown in), and have the cars pass them.
It is quite meaning less to make the cars pass tests created by the company itself.
And if you see the videos, none of the people were ever in real danger. Say, where is the test when someone crosses when the car is going to 45mph? I am not asking to put real people in real danger. But just make some robotic dummy (even something on wheels will do), and make it cross without notice infront of the car when it is doing 50+mph.
This is forward progress. Waymo has worked towards this point for years, reporting every disconnect and minor fender-bender. Now they're going to test, cautiously, in Mountain View.
As we've seen with Uber and Tesla, the "safety driver" doesn't add much safety. The "safety driver" is mostly there to drive manually in situations where the automation stops the vehicle and asks for help.
Why do they not need the safety driver? i’m sure google can afford them. i just don’t see the benefit of no safety driver. especially living in mountain view and seeing the recent tesla crash.
I'd imagine because that means no steering wheel, no manual controls, etc. All that weight can be replaced with redundant everything: engines, brakes...
> Waymo has worked towards this point for years, reporting every disconnect and minor fender-bender.
Is there a list of scenarios the Waymo car can handle?
I think there is a fundamental limitation with detecting stationary objects right in the path of the car. Did the waymo cars share this issue. If not how did they overcame it?
It's not hard. All the DARPA Grand Challenge vehicles, including ours, could do that in 2005. Watch the technical video from Chris Urmson at SXSW about how Google does self driving.
I saw couple of those videos. I didn't notice this addressed in any of those. Most of the stuff was showing how car handles intersections where it is actually slow moving..
The DARPA Grand challenge only required the cars to drive across a desert.
42 comments
[ 3.6 ms ] story [ 91.4 ms ] threadSources said that Waymo does not plan on operating its cars remotely — even in difficult situations — but it will remotely monitor them during tests. If one of the cars encounters something it doesn’t understand, such as complicated road construction, the car will contact Waymo for help recognizing the situation. After human testers give it feedback, the car will then decide how to navigate the situation.
https://www.sfchronicle.com/business/article/Exclusive-Waymo...
The most extreme type of assistance could maybe be a human drawing or modifying the vehicle's planned path (if you watch any of the Waymo example videos, you'll see the immediate path that the vehicle is planning to take). But even in this scenario, the car is still responsible for navigating that path and avoiding any obstacles.
Of course, how inconceivable that a self driving car could crash into a static barrier!
https://m.youtube.com/watch?v=6QCF8tVqM3I
[1]: https://en.wikipedia.org/wiki/Waymo#Crashes
It it could prove this 100% to itself, it should be able to prove it to the general public and the authorities in a test environment..
I think car will still be in control, you just will tell it "drive slowly here, then here"
The same as the difference between what a driver does and what a driving instructor sitting in the passenger seat does, only both remote...
Hopefully well in advance of actually hitting anything, but therein lies the problem: what does it do if it's moving at speed and figures out that it doesn't know what to do? If it continues moving, it might hit something, and if it tries to stop suddenly, it might get hit and cause a multi-car chain collision. Imagine you're following a car on a highway, in a situation where no other cars are around, a situation in which a human driver would practically never stop, and suddenly it brakes to a stop very rapidly because of some perturbation in the road that confuses it, but a human would have no trouble with. Drivers are taught to look beyond the car in front and try to anticipate what it might do; and something so contrary to expectations is going to lead to problems.
I think it would also be possible to mitigate this problem by not driving on the freeway at all for the first iterations. (Feasible if you only drive within Mountain View, and not e.g. MV <-> SF.)
I would expect a fully autonomous vehicle to constantly be planning for emergency and graceful stops, and to slow down if it doesn't have a confident plan for both types of stopping and forward progress. It should also have plans for when critical equipment fails, such as sensors or actuators as well as normal car systems like brakes, tires, engine, steering.
All others testing in California are 300 or fewer miles per disengagement.
https://www.theinformation.com/articles/inside-cruises-bumpy...
For example, make a robotic dummy walk in front of the car without notice. Does the car break and avoid a collision?
Such basic things can be formulated into a test and have the cars pass them before allowing them (even with a driver) on the roads..
It amazes me that people just stand by and allow these "tests" to happen within an unsuspecting population...
>“We’ve staged people jumping out of canvas bags or porta-potties on the side of the road, skateboarders lying on their boards, and thrown stacks of paper in front of our sensors,” Waymo said in its report.
https://www.washingtonpost.com/local/trafficandcommuting/way...
>For media day, they put on three of them. In one, a Honda convertible abruptly cut off one of the Pacificas just past an intersection, and the Pacifica braked to avoid it. In another, a car obscured by two parked vehicles suddenly backed out of a driveway. Again, the Pacifica came to a halt. The third was a faux moving-day scene, with a couch in the street and boxes tumbling into the lane, just as another car approached from the opposite direction. The Pacifica knew better than to swerve into the car. It braked until the coast was clear.
https://www.washingtonpost.com/local/trafficandcommuting/a-g...
People should not care what Waymo does.
There should be an independent body to frame these tests (with enough random stuff thrown in), and have the cars pass them.
It is quite meaning less to make the cars pass tests created by the company itself.
And if you see the videos, none of the people were ever in real danger. Say, where is the test when someone crosses when the car is going to 45mph? I am not asking to put real people in real danger. But just make some robotic dummy (even something on wheels will do), and make it cross without notice infront of the car when it is doing 50+mph.
As we've seen with Uber and Tesla, the "safety driver" doesn't add much safety. The "safety driver" is mostly there to drive manually in situations where the automation stops the vehicle and asks for help.
It also lowers barriers to an eventual widespread uber-like service. That service wouldn't be financially viable with safety drivers in each car.
Public is not asking them anything. Public does not sign up being part of a beta/alpha or what ever test, for sake of "senior management"...
Is there a list of scenarios the Waymo car can handle?
I think there is a fundamental limitation with detecting stationary objects right in the path of the car. Did the waymo cars share this issue. If not how did they overcame it?
The DARPA Grand challenge only required the cars to drive across a desert.
So I am not sure how that answers my question.