Wow, those systems are incompetent! Approaching a disabled vehicle halfway off the road at only 30 MPH they all hit the obstacle between 33% and 100% of the time. Sometimes they didn't even slow down. Nobody is driving on the interstates at 30 MPH, I'm sure the ADAS performance would be much worse at 70 or 80 MPH.

The lane keeping looks only slightly better averaging one departure every 8 miles. If I was riding with some one who ran out of their lane every eight miles I'd get out an walk. There is no horizontal scale on the following graphs so impossible to judge the frequency of the weaving, hopefully slow enough that an officer is not going to pull you over for a sobriety test. I've read some reviews in the popular press that said some systems bounce from one side of the lane to the other like ping pong balls.

The only systems that seen ready for prime time are the active cruise controls.

----------------------------------------

The Help for this program was created in Windows Help format, which depends on a feature that isn't included in this version of Windows.

Nevertheless, the marketing will continue :)

The problem with sloppy work is that the supply FAR EXCEEDS the demand

GregLocock,

I am trying to sort out a document on how LiDAR would detect objects on the road while the vehicle operated at a reasonable speed. The vehicle must not run into anything under any circumstances. I worked a lot of this out in Self Driving Uber Fatalities threads. I am satisified that a robot can avoid objects on a limited access highway while travelling at 100kph, and it can detect head-on collisions on a two-lane road at 80kph, and it can pull out from an intersection on the two-lane road with traffic approaching at 80kph. It may be annoying to drive behind this robot.

I like your term "Driver assistance". The words "Autopilot" and "Robot" should never appear on the dashboard of a car unless the robot is demonstrated to be reliable. The driver is fully responsible for safe operation of the car, and must provide their full attention to it. The robot is a back-seat driver at best. I do like the idea of robot operating controls. I can see a robot jamming on brakes just as I try to do a lane change.

--
JHG

The issue with lidar isn't the detection of objects, it's figuring out what to do about them. In the Uber accident in Arizona, the woman that was killed was detected in plenty of time to avoid her, but the collision algorithm was absurdly stupid, since no one bothered to code in any sort of tracking algorithm to determine that the woman was on a collision course with the path the car was on. All the information was collected, but the algorithm was just too dumb to do anything with the information it had.

TTFN (ta ta for now)
I can do absolutely anything. I'm an expert! https://www.youtube.com/watch?v=BKorP55Aqvg
FAQ731-376: Eng-Tips.com Forum Policies forum1529: Translation Assistance for Engineers Entire Forum list http://www.eng-tips.com/forumlist.cfm

IRstuff,

The problem in Utah was that all the devices that were to take evasive action, including the driver, were turned off.

--
JHG

No doubt. But in Arizona, the driver wasn't alerted until the collision was unavoidable, although the car had the information of the potential imminent collision for about 5 seconds before it finally gave warning.

TTFN (ta ta for now)
I can do absolutely anything. I'm an expert! https://www.youtube.com/watch?v=BKorP55Aqvg
FAQ731-376: Eng-Tips.com Forum Policies forum1529: Translation Assistance for Engineers Entire Forum list http://www.eng-tips.com/forumlist.cfm

The computer pretty much ignored the odd object heading into it's path until it figured out what it was.

Actually, it knew what the object was, even to extent of detecting the bicycle the woman was pushing. But, that's kind of my point, regardless of what the object was, and whether the detections were consistent, SOMETHING was on a potential collision course, that should have been enough for an SAE level 1 or 2 system to immediately demand attention from the driver.

TTFN (ta ta for now)
I can do absolutely anything. I'm an expert! https://www.youtube.com/watch?v=BKorP55Aqvg
FAQ731-376: Eng-Tips.com Forum Policies forum1529: Translation Assistance for Engineers Entire Forum list http://www.eng-tips.com/forumlist.cfm

IRstuff,

It is nice that it identified the object, but that is irrelevant. There is no object a robot car can be allowed to run into. In three LiDAR scans, the robot knows where the object is, its velocity and its acceleration. It should start taking evasive action immediately.

--
JHG

I never implied otherwise, but if it's Level 1, it is not capable of performing all dynamic driving tasks and requires the driver to complete tasks that are outside of its operational design domain (ODD)

TTFN (ta ta for now)
I can do absolutely anything. I'm an expert! https://www.youtube.com/watch?v=BKorP55Aqvg
FAQ731-376: Eng-Tips.com Forum Policies forum1529: Translation Assistance for Engineers Entire Forum list http://www.eng-tips.com/forumlist.cfm

I use the adaptive cruise control a lot and although it takes a leap of faith the first time you use it it will brake HARD if it needs to.

When not in cruise it will sound a warning and light up the dash if it thinks you're close to hitting something - saved me a couple of times already in a year and a half.

Is it fool proof - no. If you're following a car which moves into the central median to turn right( left in the US) it can think you're about to hit it as it's still in the radar view and start braking hard for no reason risking a rear end collision.

Also can't see further than the car in front.

To avoid false readings the radar view is quite narrow beam so someone changing lanes in front of you doesn't get picked up until they've moved lanes.

But still a great piece of technology which works really well.

I really don't like the keep lanes technology though - Really can't see the point. If you've got to steer then you should be able to keep the car in a straight line or within your lane on a junction.
I turn it off all the time.

Remember - More details = better answers
Also: If you get a response it's polite to respond to it.

Quote (drawoh)

There is no object a robot car can be allowed to run into.

What about a piece of paper blowing in the wind?

What about a garbage bag? Do you know what's inside it?

What about a squirrel running across the road? Cat? Dog? Small deer? Moose? What's the threshold? Is a human infant within that threshold?

What about a bouncing ball? The ball itself probably is not a significant risk factor. What's chasing the ball ... might be. On a similar note ... a baby black bear may be followed across the road by a huge momma bear. Does the automation system know that?

Does the action that the robot driver take, factor in what's in the rear-view mirror? Avoid a squirrel, get run over by a dump truck.

What's the threshold for doing something illegal? It is always illegal to go through a red traffic signal without stopping. What about that cement truck filling your rear-view mirror?

This particular human driver *always* shoots a glance in the rear view mirror upon a situation arising that could warrant anything more strenuous than a normal stop. (actually, in reality, the action to be taken upon a traffic signal changing state to yellow will already be planned based on what's behind me even before the signal changes state)

What about traffic entering a motorway lane from a merge lane? Is the automation smart enough to realise that the lane that the merging traffic is in, is about to end, and that even if that merging traffic is not yet in an interference condition, it soon will be? This particular human driver spots that situation developing, and takes preventive action before an interference condition develops - be it changing to the next lane over, or subtly slowing down or speeding up to align my own vehicle with a gap in the merging traffic so that the merging traffic can merge smoothly without anyone having to do anything evasive or suddenly.

I have my doubts!

I like your point about mirrors. The BMW system hit a peak deceleration in excess of 1g when following a car decelerating at 0.6g.

Cheers

Greg Locock


New here? Try reading these, they might help FAQ731-376: Eng-Tips.com Forum Policies http://eng-tips.com/market.cfm?

Quote:

Actually, it knew what the object was, even to extent of detecting the bicycle the woman was pushing.

Sure, but it pretty much waffled about trying to decide what she and the bike were until it was about to hit her before the final determination was made.

Quote:

This particular human driver *always* shoots a glance in the rear view mirror upon a situation arising that could warrant anything more strenuous than a normal stop.

A good driver should know what's behind them before a situation arises.

Not too long ago I changed lanes when I panic brakes one night for a family of skunks because I already knew there was a transport not far behind me.

I find that the pilot assist and adaptive cruise control are very helpful to maintain driver alertness on a long trip because you can move around, stretch and adjust your position to keep from becoming frozen in place. It also helps when the kids scream and you need to look back to make sure they are not killing each other.

Quote:

Sure, but it pretty much waffled about trying to decide what she and the bike were until it was about to hit her before the final determination was made.

Sure, but at no time did it not detect a target that could be damaged or damage the car. Moreover, the fact that it couldn't consistently classify the target is irrelevant, since there was always a target present; the fact that the system was designed to account for misclassifications is a design flaw. The fact that the system was designed to brake when the stopping distance was already too long is a design flaw.

I've been down this road with other design engineers who were so pleased with their target processing that they failed to account for why the system exists in the first place, which is to prevent accidents.

TTFN (ta ta for now)
I can do absolutely anything. I'm an expert! https://www.youtube.com/watch?v=BKorP55Aqvg
FAQ731-376: Eng-Tips.com Forum Policies forum1529: Translation Assistance for Engineers Entire Forum list http://www.eng-tips.com/forumlist.cfm

Someone on this forum put it best by saying that these systems should be looking for clear road.

Of course, looking for potential targets and conflicts and predicting conflicts is part of that, but the objective should still be to predict the clear path between the potential conflicts.

It's akin to "you go where you look, so look where you want to go" ... look at the clear space between the various obstructions, and go there.

Which is exactly what I put in my first post - it ignored the object until it determined what it was. Too bad by that time it was too late to avoid her. The accident made it obvious the system was flawed.

As for the unknown object being damaged or damaging the car. The computer had no idea what the object was so how could it determine if damage was possible? That kind of expectation falls into the "what objects need to be ignored" category. Otherwise, a truly self driving car will get stuck at an object it should be driving past. This part is much harder than simply stopping at every possible hazard but is something that eventually needs to be figured out if cars are going to truly drive themselves without human oversight.

The Tesla that went under the truck tried for clear road by shooting between the object that just crossed it's path and the object that was headed into it's path.

My '18 Yukon has lane keeping warning symbols, too small and too far below my line of sight to really notice, and irregularly active.

It also has 'adaptive' cruise control, which enables you to set one of three following distances.
I can't make it adapt as documented.
I think it's triggered BIG TIME just once. I don't know what set it off.

I'd consider each a work in progress.

The Yukon also has a modern version of the old Guide automatic dimmer, which works flawlessly. ... after you spend a month reading the book and experimenting to figure out how to make it start working.

The Human interface is grossly inferior to that of my 2010 Mountaineer, as was my 2013 Navigator.
I think Ford closed the wrong division.



Mike Halloran
Corinth, NY, USA

Quote (BrianPetersen)

What about a piece of paper blowing in the wind?
...

I worked this out roughly somewhere in thread815-436809: Self Driving Uber Fatality - Thread I. Since then, I have done a detailed calculation which I have just posted to my website. This is preliminary and I am not sure what to do with it. It does show that a robot on a highway at 100kph, can detect an object 104m in front, and decelerate to a halt at a rate such that a reasonably attentive, responsible driver in a reasonably functional vehicle behind them, can decelerate, and not hit them. The object is viewed as a raster scan with a resolution of 180mm. The robot can see small objects, but it probably cannot identify them.

Your assumption is that there is a list of objects we can safely decide to run into. I claim that the robot's ability to identify objects is not sufficient or reliable enough to make this decision. No thing can be run into safely by a robot.

Note how at 104m and 100kph, the robot must start taking evasive action. On my robot, the forward LiDAR actually has a range of 250m, but out there, the resolution is around 440mm.

I too check my rear view mirror when I hit the brakes. There have been a number of horrific car accidents where fully halted vehicle was rear-ended by drivers who made no attempt to slow down. They were asleep at the wheel or something. The accident is their fault, and they should be charged with dangerous driving.

--
JHG

Quote:

As for the unknown object being damaged or damaging the car. The computer had no idea what the object was so how could it determine if damage was possible?

Again, the Arizona Uber knew, more than once, the object detected was not allowed to be hit.

Quote (https://www.ntsb.gov/investigations/AccidentReport... )

It initially classified the pedestrian as a vehicle, and subsequently also as an unknown object and a bicyclist.

The clear fatal flaw was that every time the classification changed, it drops the track history. The initial vehicle detected suddenly disappears, but an unknown target appears in about the same place 0.4 seconds later and approximate 1.5 ft away; the next detection 1 second after that was vehicle again, but track history dropped again. That is just plain stupid, since we've know since the 60s not to do that in ICBM warhead tracking; it's illogical for physical objects to disappear and reappear, and trackers have been developed to deal with those sorts of scenarios, decades ago.

TTFN (ta ta for now)
I can do absolutely anything. I'm an expert! https://www.youtube.com/watch?v=BKorP55Aqvg
FAQ731-376: Eng-Tips.com Forum Policies forum1529: Translation Assistance for Engineers Entire Forum list http://www.eng-tips.com/forumlist.cfm

This has popped up on Slashdot.

Cory Doctorow: 'Self-Driving Cars are BS'

The article is classified under "entertainment". He seems to be addressing issues other than robot control. I have argued in other threads that autonomous robot cars will be a service, not a possession, which negates some other issues he has here.

Interesting stuff though.

--
JHG

Quote (drawoh)

The article is classified under "entertainment". He seems to be addressing issues other than robot control. I have argued in other threads that autonomous robot cars will be a service, not a possession, which negates some other issues he has here.

Congestion is the killer in the cities. Large scale private ownership of self driving cars is unmanageable. Their use would require heavy regulation.

The only thing that currently makes congestion manageable is the time cost of the driver. Remove that and the tragedy of the commons would be extreme.