GM's Cruise revises self-driving software after San Francisco crash General Motors autonomous unit Cruise has issued a Safety Recall report, pulling software that governed how its AVs behave when making an unprotected left turn after one of the vehicles was involved in a crash. An unprotected left turn is when you want to turn at an intersection where oncoming traffic (either turning right or …
You swerve out of the way of the wrong lane/speeding car and mow down a bus stop full of orphans, while the other cars drives off, it's you who get the naughty step.
Let the other car hit you (assuming that you are in the massive pickup and the oncoming car is a prius) and then sue them for (insert little finger) $1M
Having recently taught my kids how to drive, I became much more aware of how i rely on very subtle visual cues to predict how other drivers will behave. So personally, I can't see how self-driving software can achieve such a high level of understanding at least for a number of years.
The money behind these companies will push through all of the concerns for self-driving cars with governments and we will end up with lots of self driving cars. Then in the end, it will be the human drivers who will start to learn the faults and limitations of these cars and give them a wide berth. Maybe we just stick a big L plate on all of these cars so we can know that they can't be trusted. I think making these self driving cars much more visible is the safest approach rather than relying on AI to understand human drivers and pedestrians.
I would think that the main thrust of development of these self-driving cars is to make them drive as predictably as possible.
TBH, relying on subtle clues from the other meat-based driver is probably not the best general scenario for safe driving, although I know what you mean: as an experienced driver you get a feel for what the other drivers are going to do in all sorts of situations. I was always taught that the safest drivers are the most predicable ones and I try to make sure that my intentions are pretty unambiguous. Honestly, given some of the drivers that I see on the road, I'm astounded that there are not *more* accidents than there actually are.
Personally, I think that after the wrinkles have been ironed out of the current vehicles, we are going to find out that the vast majority of accidents involving them are going to be caused by irrational, erratic or downright illegal behaviour by human drivers, if that isn't indeed already the case. That certainly seems to be the case in the incident reported on in the article.
"He added: "We hear from a lot of women who say they feel a lot safer at night riding home in a car without someone in it that they don't know.""
And when the women realise they don't know who might be able to control or divert the driver-less car, or if it suddenly fails to proceed / chooses a not so well-known route and an operator is being sent to sort it out ?? How safe will they feel then ?
They should feel safer! Full audio and video recordings of them being picked up alone... dropped off alone, with the added benefit of their locations and which times they're alone being uploaded to a large real-time network and database. All this without having to worry about another human being you'll see face to face, just truly alone! Safety FIRST people... safety first!
If you really think about anyone rolling around alone in a vehicle they can't control... well they're just 1 ski-mask away from a memorable night. At least if they're driving the vehicle themselves they could speed away but, with a robot... a robot will simply feed you to the criminal.
I don't know the exact junction layout, but I get the impression that it is unlikely that a Prius could make the turn at 40mph. This should have been enough to tell the Cruise AV that it wasn't turning or to make the turn quickly, keep right and be prepared for an out-of-control Prius to enter its path. Or better still, don't even attempt the turn.
If too much for an AV then it's not ready for real-world driving.
We know from the report that the Prius was screaming down the right-turn-only lane at 40mph. If I were in the position of the Cruise and has seen the approaching car, I probably would have stopped to see what was going to happen. As you rightly surmise, if the vehicle was going that fast, unless it braked extremely hard, it wasn't going to be making that turn.
It is going to be interesting analysing some of these scenarios after the fact, since in at least some cases, these automatic vehicles are often going to have more information available then the average driver and so might make different decisions that we would.
We have already seen situations where a self-driving car detected a threat that a human would never have been able to see, like where lidar detected a cyclist that wasn't visible to human eyes and was therefore able to prevent an accident. In this incident, the Cruise knew that the Prius was in a right-only turn lane: perhaps other drivers in the intersection would not know this.
Yeah in that situation with a car speeding down the right hand turn lane I probably wouldn't complete my left turn either. It is sort of a damned if you do damned if you don't situation - the kind of thing that causes accidents that are impossible to avoid without telepathy to determine another driver's actual intentions (and that assumes they didn't have the intention of turning right and then realized at the last second "oh shit this isn't the street I want to turn right on and go straight through in a right turn only lane)
Really? In that situation the decision is a, don't make the turn or, if already committed then b, floor it to get through the intersection. The robocar did the unacceptable option of c, do nothing. Flooring it might have cleared the intersection and avoided the accident but stopping in front of a car going too fast was the wrong decision. Sure, legally the Prius might have been speeding but the robotaxi has the higher level of fault for turning left in front of oncoming traffic without enough time to clear the intersection.
> "lidar detected a cyclist that wasn't visible to human eyes" -- just curious what scenario that is.
I don't remember the exact details but it was reported here. IIRC the cyclist was to the right on the other side of a hedge. So either the Lidar was high enough to see over the hedge or could see through it.
My guess is that the robocar did not get a good fix on the Prius's distance, heading and speed before beginning the turn manouver.
This raises the question of how far away the robocar is able to accurately track other vehicles, since self driving cars are touted as having 360 degree situational awareness and accurate sensors in order to outdri e stupid meatbags.
Why was the Prius, obviously speeding, not identified as threat before the robocar began its turn? Why did the robocar believe that a vehicle which was already behaving in a dangerous way would definitely obey the lane rules and make a turn?
And that was my biggest surprise when I started trying to use Tesla's FSD - how poor it was (or choose not to) look far enough ahead to anticipate and react to what's coming. And that also explains why it drives like a 16-year-old. One of things I constantly had to remind my daughter when she was learning to drive was to look way beyond the car in front of you.
Key paragraph:
"...NHTSA...claimed [NOT proved] the oncoming vehicle was...estimated at [NOT measured]...40mph in a 25mph zone...in the right-turn/bus-only lane as the Cruise AV was attempting to make its unprotected left turn."
Who cares what the idiot in the Pious clown car was doing? GM is 100% responsible for its clown-car deathtrap.
Here is what a traffic-court judge would tell you:
- You are not owed a left turn.
- Your left turn is under your control.
- Do not turn unless conditions are unambiguously safe.
- PERIOD.
A proper judge is not going to entertain teenager theories of "AV" loons.
I hope that both CA and GM are sued until their bad ideas end up in the grave along with the people they kill.
They say that they "disabled unprotected left turns" - in a self-driving car, did I get that right? So, "Hey, car, drive me to the airport." - "I'm sorry, Dave, I'm afraid I cannot do that." Or will it make three right turns around the block every time instead now?
Comments above about subtle cues from other meat-based drivers, not being visible due to dark screens etc: I'm not sure it's the actual driver you'd get subtle cues from, more the behaviour of the vehicle under the meat's control. However, I would absolutely agree that as your driving experience increases, your natural wariness increases, in the same way they say that 2/3rds of verbal communication is actually body language (provided you're not a driver in the latest go-faster make of machine - was originally BMW drivers, then Audi, I suspect Tesla now high on the list). Just yesterday I found myself slowing just because I felt the road arrangement I was entering had potentially hidden dangers. How AI/driverless is ever going to get to that sort of awareness, I don't know.
The Register 
Biting the hand that feeds IT
Copyright. All rights reserved © 1998–2024
