As I was sitting in the car playing with the new autopilot, I noticed most of the disengagements are due to other drivers pulling dumb moves. For example, they will try to merge at an unsafe speed or do things that are so unpredictable that autopilot really doesn't have a way to react to it until it is too late. Just how are we going to program all of these things into the AP computer? Is AI really so smart that it knows a guy is going to change lane without using blinker or is it going to predict what the other driver is going to do based on the cars in front of them? I honestly don't know how this is going to happen within the next couple years as there are so many variables? I think other driver is probably a bigger problem than poor road conditions (lack of lines, constructions, etc)
Can Level 5 self driving really happen with other drivers?
- Thread starter mattack4000
- Start date
You can't fix dumb. AP will not be able to prevent all accidents or protect against all scenarios (while sill being usable).
That said, I expect the AP code to use blinkers and such as a secondary data point to the physical motion of other vehicles. I commute through a double lane roundabout twice a day, and would have replaced my car 3+ times if I wasn't paying attention to other vehicles' speed and rotation.
Exactly, so how are we going to trust the computer bring us from point A to B. I just had this in my head as I keep having to turn off AP due to other drivers.
I don't even think blinker is enough, there has to be more. Here is another example. I was coming to a merge at 25mph. The woman next to me never really moved behind the car at the merge, she stayed to her side for the most part. AP took off to catch up to the car in front trying to close the gap. If I didn't hit the brakes, the other driver would likely just swerve into me thinking I should have let her go. How do you fix that? I don't even think blinkers would help in this situation as the woman didn't use any.
I don't even think blinker is enough, there has to be more. Here is another example. I was coming to a merge at 25mph. The woman next to me never really moved behind the car at the merge, she stayed to her side for the most part. AP took off to catch up to the car in front trying to close the gap. If I didn't hit the brakes, the other driver would likely just swerve into me thinking I should have let her go. How do you fix that? I don't even think blinkers would help in this situation as the woman didn't use any.
Right, I'm saying it shouldn't use blinker/ lack of blinker as a primary data point.
I'm not clear on the the position of everyone in the situation you mention, but AP can't stop someone from swerving/ merging into you. Best it can ever do is avoid pinch points based on merges/ monitoring other's driving style (or lack thereof) and such.
On the roundabout, I have people cutting me off, pulling out into me, changing lanes (not allowed), turning into me. Great training ground for defensive driving...
nwdiver
Well-Known Member
AI is how it's 'fixed'; So long as humans are in the mix AP will never be perfect but it doesn't need to be... as Musk has said it just needs to be ~10x better than a human driver which is a really... really... REALLY low bar. Your experiences with the current AP are not what L5 will be. It's a process and AP learns. As you said, 'You hit the brakes' AP will learn that too...
I think there is a misunderstanding. I'm saying the computer will predict pinch points and will observe vehicle motion for actual trends.
I'm also saying AP can't avoid bad drivers entirely.
I'd like it to be better than me. 
Right now, EAP is mostly smart speed control and stay-in-lane and does not have threat detection/ object prediction.
It will be interesting to see how much FSD will bend normal driving rules (use of shoulder, speed limit) to avoid collisions.
Right now, EAP is mostly smart speed control and stay-in-lane and does not have threat detection/ object prediction.
It will be interesting to see how much FSD will bend normal driving rules (use of shoulder, speed limit) to avoid collisions.
One of the biggest frustrations I have with driving right now is the divide between those who drive dangerously cautious to those who drive recklessly. How they drive is dependent not on the rules of the road, but on their own personality/style.
To a large degree driving is a social activity. It's full of being aware of others, and negotiations between people. It's having instincts about what someone is going to do before they do it.
How can you teach a machine to drive safely in a world where it's blind to both of those thing?
The answer to that is you likely don't. That's not useful information to it long term.
Lots of drivers are on the road that wouldn't be there if we had autonomous cars. As we have more ways of getting from Point A to Point B means we can toughen the requirements for earning a licence, and increasing the requirements for keeping the license. I don't see any other way of combating the level of disregard for other people's lives that drivers on the road currently have. It's so bad that fatalities are up in recent years even though we have safer cars than ever before.
We also have to fix an infrastructure problem where we have roads that are absolutely dangerous whether you're a human driver or autonomous driver.
We won't see full level 5 self-driving cars suddenly appear. Instead we're going to see autonomous driving services like waymo has over more areas where certain roads are blacklisted. As time goes on more, and more roads will be allowed.
Accidents will happen and we'll adapt as we go. The promise of autonomous cars is you only have to teach it once. Where with a human you have to teach all the humans.
At some point it's going to come down to how tolerant people are to deaths as a result of autonomous driving. Some people won't accept even not at fault fatalities (as if the car could do magic), and others are of the "as long as it's just better than a human mindset. I'm somewhere in the middle where I expect a >5 times safer on average from an autonomous car. I do because I think there is a lot of low hanging fruit that won't be grabbed if we make it too lenient.
To a large degree driving is a social activity. It's full of being aware of others, and negotiations between people. It's having instincts about what someone is going to do before they do it.
How can you teach a machine to drive safely in a world where it's blind to both of those thing?
The answer to that is you likely don't. That's not useful information to it long term.
Lots of drivers are on the road that wouldn't be there if we had autonomous cars. As we have more ways of getting from Point A to Point B means we can toughen the requirements for earning a licence, and increasing the requirements for keeping the license. I don't see any other way of combating the level of disregard for other people's lives that drivers on the road currently have. It's so bad that fatalities are up in recent years even though we have safer cars than ever before.
We also have to fix an infrastructure problem where we have roads that are absolutely dangerous whether you're a human driver or autonomous driver.
We won't see full level 5 self-driving cars suddenly appear. Instead we're going to see autonomous driving services like waymo has over more areas where certain roads are blacklisted. As time goes on more, and more roads will be allowed.
Accidents will happen and we'll adapt as we go. The promise of autonomous cars is you only have to teach it once. Where with a human you have to teach all the humans.
At some point it's going to come down to how tolerant people are to deaths as a result of autonomous driving. Some people won't accept even not at fault fatalities (as if the car could do magic), and others are of the "as long as it's just better than a human mindset. I'm somewhere in the middle where I expect a >5 times safer on average from an autonomous car. I do because I think there is a lot of low hanging fruit that won't be grabbed if we make it too lenient.
I like your post. Could you expound on what you think will be the cause of at-fault crashes resulting death for AP? It seems to me that would require failure to:
- maintain lane
- recognize object in path
- recognize traffic device
- exercise due care (not giving bicyclist, pedestrian... sufficient space to avoid collision)
- keep working (hardware failure during operation)
nwdiver
Well-Known Member
It's even better than that. Every AP car will learn from every mistake every AP car makes. It's an exponential growth curve of learning.
GatorGuy
Member
I doubt that a computer will able to predict human driving behavior than other humans for a while, but it can observe and respond to it better than we can. Unlike a human, the computer will know what is around it and coming towards it at every possible second unlike any human. A computer can also respond quicker than a human can and can make better decisions because it will have more information. Where as a human driver might hit the brakes to avoid an accident, it could decide to swerve into another lane that is empty, etc.
Try to follow traffic law.
Don't hit things.
It's worth remembering that a lot of collisions are caused by a failure to react in time. This happens because meatsacks are reckless, and take time to react. Simply by not being reckless, a human or L5 system can reduce collisions significantly. A human can see small signs that allow them to anticipate and be even safer. An L5 system need not anticipate to be extremely safe since it could make up for lack of anticipation with superhuman vision and reactions.
Don't hit things.
It's worth remembering that a lot of collisions are caused by a failure to react in time. This happens because meatsacks are reckless, and take time to react. Simply by not being reckless, a human or L5 system can reduce collisions significantly. A human can see small signs that allow them to anticipate and be even safer. An L5 system need not anticipate to be extremely safe since it could make up for lack of anticipation with superhuman vision and reactions.
It seems like the rear facing repeater camera is perfectly positioned to handle that. It is as far to the side as possible (other than if it were mounted at edge of side mirror) so has the best view down the gap between cars. Ultrasonic would not be used for this situation.
30 MPH is 44 ft/sec so if the car does not change lanes abruptly, the motorcycle should see the maneuver and the car should see them in time to avoid colliding. Quite similar to a normal highway merge or lane change. Perhaps even easier than a left turn.
sillydriver
Member
I think this is a great statement of one of the two things it will be very hard to program an automated system to deal with, the other being hand signals / body language from police or other personnel directing traffic.
I really like this post. I have to constantly evaluate other drivers for their intentions and to predict what they will do when they close enough to me. You've seen them...
The driver who is passing in the right lane to gain a one car position in the left lane. You know that driver will cut in front of you (in the left lane) when they have just enough space. You saw them in the rear view mirror 10 car lengths back.
The driver who is in the left lane going 5 under the limit. You see them 1/4 mile ahead. You have to make the decision now to move right to go around.
The truck that is gaining at 1 mph on the truck in front of him while you approach in the left lane. You know they will simply pull out regardless of the speed differential between you and them.
Until all drivers can be predictable by an AI (i.e. all drivers are AI) these situations will continue.
Schumpeter
MS 90D owner. EV and renewable energy enthusiast!
Similar threads
