Can cars ever really FSD until they can read body language?

[WillyC]

The more I use autopilot the more I see the complexities of self-driving cars. Here's an example that would be impossible for a car to deal with correctly unless it can read human body language.

I was on autopilot in a 50/60 limit with a very wide lane, the car was in the middle of the lane as you'd expect with a lot of space either side. Up ahead I could see a woman standing at the edge of the road looking up and down the road waiting to cross. When she was 10-15m in front of the car she stepped into the road, clearly with the intention of crossing quickly behind me. As you'd expect, the car sees a pedestrian in its lane and the alarm goes off and it slams on the brakes. This kind of edge case is extremely difficult to deal with.

A much more common problem will be pedestrian crossings. Are the two people standing there waiting to cross or just having a chat? A human would instinctively know this by looking at what the pedestrian is paying attention to, but getting neural networks to understand and interpret this situation correctly is much more difficult.

Just food for thought.

 

[pgkevet]

That is something Musk touched on with the HW3 launch. ..Claimed that in time the cars would read expression and body language. ... at least he didn't give a time scale for it to happen...

 

[WarpedOne]

You mean body language of a dog or a cat?

 

[Glan gluaisne]

Should be possible to determine intention from movement at least as well as humans can, I'd have thought. We've learned to recognise subtle cues from the way someone moves to be able to make a pretty good guess as to what they are most likely to do. It's probably one of those things that the system will take some time to learn, much as it is for humans, but there's no reason why, in theory, it shouldn't be able to work at least as well as it does for us. My guess is that initially there will be an abundance of caution, with the car slowing, or stopping, when it may not need to.

 

[pgkevet]

You mean body language of a dog or a cat?

I'd guess that priority would apply to guide dogs after people learning. Perhaps the car will look out for large ginger cats carrying lasagne too...

 

[MrBadger]

The more I use autopilot the more I see the complexities of self-driving cars. Here's an example that would be impossible for a car to deal with correctly unless it can read human body language.

With the right amount of learning, AI is extremely good at picking out nuances. Whilst its not going to be looking for particular scenarios, the nuances from different actions, be it body language, car behaviour, road conditions etc will ultimately be part of the neural network that will influence the cars response to different human and vehicle behaviours.

tl;dr, response to these behaviours will ultimately fall out in the wash...

 

[Jason71]

The more I use autopilot the more I see the complexities of self-driving cars. Here's an example that would be impossible for a car to deal with correctly unless it can read human body language.

I was on autopilot in a 50/60 limit with a very wide lane, the car was in the middle of the lane as you'd expect with a lot of space either side. Up ahead I could see a woman standing at the edge of the road looking up and down the road waiting to cross. When she was 10-15m in front of the car she stepped into the road, clearly with the intention of crossing quickly behind me. As you'd expect, the car sees a pedestrian in its lane and the alarm goes off and it slams on the brakes. This kind of edge case is extremely difficult to deal with.

A much more common problem will be pedestrian crossings. Are the two people standing there waiting to cross or just having a chat? A human would instinctively know this by looking at what the pedestrian is paying attention to, but getting neural networks to understand and interpret this situation correctly is much more difficult.

Just food for thought.

¯\_(ツ)_/¯

 

[Js1977]

Im with the first poster. I consider myself very good at spotting people on the road that are likely to cut me up or misbehave. Sometimes its i can tell just by the car they are driving without looking at them. It ranges from learner drivers to OAP's that never even took a test, boy racers etc.

And then there is the other side of me that says never say never and It most probably will be possible (where there is a will there is a way). This is also the bit where i get my tin hat on because we are talking about perception and intuition here in a machine that will eventually keep learning on its own and eventually.....

 

[ItsNotAboutTheMoney]

The more I use autopilot the more I see the complexities of self-driving cars. Here's an example that would be impossible for a car to deal with correctly unless it can read human body language.

I was on autopilot in a 50/60 limit with a very wide lane, the car was in the middle of the lane as you'd expect with a lot of space either side. Up ahead I could see a woman standing at the edge of the road looking up and down the road waiting to cross. When she was 10-15m in front of the car she stepped into the road, clearly with the intention of crossing quickly behind me. As you'd expect, the car sees a pedestrian in its lane and the alarm goes off and it slams on the brakes. This kind of edge case is extremely difficult to deal with.

A much more common problem will be pedestrian crossings. Are the two people standing there waiting to cross or just having a chat? A human would instinctively know this by looking at what the pedestrian is paying attention to, but getting neural networks to understand and interpret this situation correctly is much more difficult.

Just food for thought.

The bigger problem isn't dumb pedestrians, it's smart pedestrian who know that the AV has seen them and will stop. Same will apply to drivers.

That's why I believe that autonomous vehicles will actively record and report bad actions to the authorities. It'll be the only way to stop it.

A good AV will perform path analysis.

 

[GRiLLA]

If you can learn to do this as a human, it's something that you can teach a neural net with enough input. Also while we may feel we have superior ability we only have one pair of blinking eyes and a fairly easily overloaded CPU, the evidence of which is the number of accidents caused by humans.

 

[Glan gluaisne]

The bigger problem isn't dumb pedestrians, it's smart pedestrian who know that the AV has seen them and will stop. Same will apply to drivers.

That's why I believe that autonomous vehicles will actively record and report bad actions to the authorities. It'll be the only way to stop it.

A good AV will perform path analysis.

Been a problem for years in warehouses running driverless pallet trucks etc. People quickly worked out that these things would always stop if they just walked in front of them, and it caused problems when it became second nature for everyone sharing a workplace with these things to just make them stop because they could. The solution in warehouses has been to segregate the areas where people work from the areas where semi-autonomous vehicles work, not something that will work on public highways.

I'm convinced that it won't be a hard thing for any self-learning system to do, though. If humans can learn to do it, in the time it takes them to learn to drive, then I'd guess it should be easy enough for a self-learning system to do. I think there are much harder challenges for self-driving, the sort of things that even humans find challenging, that draw upon human attributes that haven't been acquired as a result of driving a car.

 

[Irata]

I doubt the camera resolution is good enough to read facial expressions and subtle body language from the braking distance needed. It's far lower resolution than the human eye. So it will likely always act on the side of caution, AI can't help if the input data isn't available.

 

[Glan gluaisne]

I doubt the camera resolution is good enough to read facial expressions and subtle body language from the braking distance needed. It's far lower resolution than the human eye. So it will likely always act on the side of caution, AI can't help if the input data isn't available.

I'm not sure it needs to, as I suspect that the main cues come from motion vectors from things like knees, feet, hands etc, and they way these change in 3D space. How well can a human driver see facial expressions at the typical thinking + braking distance of ~23m at 30mph? Not that well, I suspect, especially in just a short glance away from the road ahead.

 

[Irata]

I'm not sure it needs to, as I suspect that the main cues come from motion vectors from things like knees, feet, hands etc, and they way these change in 3D space. How well can a human driver see facial expressions at the typical thinking + braking distance of ~23m at 30mph? Not that well, I suspect, especially in just a short glance away from the road ahead.

Agreed... I wouldn't call that body language, I would call that intent based upon motion.

The problem suggested was people waiting near a crossing.

 

[Jason71]

Agreed... I wouldn't call that body language, I would call that intent based upon motion.

The problem suggested was people waiting near a crossing.

Of course technically there is no requirement for the car to stop unless the people have actually stepped onto the crossing so determining intent is not actually needed in that instance

 

[WillyC]

The bigger problem isn't dumb pedestrians, it's smart pedestrian who know that the AV has seen them and will stop. Same will apply to drivers.

That's why I believe that autonomous vehicles will actively record and report bad actions to the authorities. It'll be the only way to stop it.

A good AV will perform path analysis.

If the car intentionally hits 1/1000 pedestrians that walk out in front of it, that'll pretty soon put an end to it!

 

[HVM]

This is not a problem, Neuralink will install FSD to pedestrians...

 

[pgkevet]

If the car intentionally hits 1/1000 pedestrians that walk out in front of it, that'll pretty soon put an end to it!

Letting AI learn without parameters it;d start by hitting all pedestrians in the road until it gets soem negative feedback on it's actions.

AI is out there, it can't be bargained with, it can;t be reasoned with, it doesn't feel pity or remorse or fear, and it absolutely will not stop... EVER, untill you are dead!

 

[Spacep0d]

I spent 20+ years as a motorcycle rider. One learns to predict behavior exceedingly well because one's life depends on it!

One obvious example of predictable human behavior is the tailgater. A tailgater wants to do something they're being blocked from doing (excluding road rage). So, if someone is tailgating to the side, they will probably use any available space in front of you if it opens up, more so the longer they're trapped and the more aggressive the driver's profile is. Pass with caution or invite them in front. People being held up by slow pokes often make big, sudden, sweepingly dangerous moves to get around said slowpoke, which is why slowpokes are far more dangerous than they realize (no matter what their excuse or ignorance).

My guess is that FSD would account for tailgaters (or should), whether they're behind, to the side, etc. Preventive behavior as good drivers is even more important than reactive behavior.

This also dovetails in to why speed limits are set the way they are on American surface streets. When the speed limit is too low (below the 85th percentile based on an engineering and traffic survey), this creates more lane-changes and more accidents. Too slow is dangerous. Too fast is too, obviously, which is why the 85th percentile is used. Speed parity is what's important, which is why the Germans have the Autobahn and fewer accidents per driven mile (and the superior driver training contributes to this).

The good thing about FSD here is that even without reading body language, its 'eyes' are *always on* and in every important direction. This is far superior to many unskilled or low-effort human drivers. Some human drivers hardly look in any direction or the correct direction long enough because they have lapses in judgement, poor judgement, are distracted, etc.

Even without the nuanced skills of good drivers, FSD will be better than most drivers soon enough if only it can correct phantom braking and get better on B-roads, unmarked roads, cities, etc.