[RahulNagaraj]

While Autonomous Vehicle (AV) startups are raising billions of dollars toward developing 'true self-driving' cars, experts and industry executives believe that human supervision may permanently be needed to get robot cars out of trouble.



According to reports, making robot cars drive safer than people is immensely difficult as the software lacks a human's ability to predict and assess risk quickly, especially during unexpected incidents or "edge cases".

When asked if he could see a future where remote human overseers could be removed from operation, Kyle Vogt, CEO of Cruise, a unit of General Motors, acknowledged that he could provide his customers peace of mind, knowing that there is always a driver if needed.

As per reports, most autonomous vehicles on public roads use humans in remote control centres as backups to assist them in edge cases. This is where the autonomous tech piloting the car gets confused, even by mundane situations like a child suddenly running onto the street or an object in the middle of the road. The remote humans then intervene to get the car out of these situations.

Reports also state that 'truly autonomous' vehicles are far behind the optimistic rollout schedules predicted a few years ago.

Source: Reuters

Link to Team-BHP news

[Sk8r]

So to sum up, they can never be something called " Truly Autonomous". Will be interesting to see where all those billions of dollars have gone !

[GForceEnjoyer]

Honestly, this is something that should be realised by more people. Unless you equip cars with supercomputers running code that can fully replicate the human thought process in such split-second decision making, there will never be "fully-autonomous" cars. Sure, existing technology will get better, but the "utopian vision" of someone getting into a driverless car and having it navigate city traffic and then the highway to work is a very unlikely thing. It's just as the article states: the technology can greatly assist drivers, sure. But it won't be able to replace them for a very long time. Just my two cents; information on the technical side of this matter would be welcome.

On a lighter note, I will have an insane amount of respect for the person who comes up with a system that can safely handle cattle running into the middle of the road, while a rash driver is overtaking at three times the speed limit from the wrong side in a car with bald tyres and non-functional lights, while simultaneously trying to avoid potholes and unmarked speed breakers that would end up totalling the car. All while there are no lane markings and road signs and there is debris lying on the road from "patchwork" carried out two years ago.

[androdev]

We will be forced to accept some collateral damage for the greater good. Making people lazy has always been profitable and profitability always triumphs. Finally, we Indians might appreciate the blessing in disguise of having such "unscientifically designed" roads.

[aghate]

This is a very good video from Ryan at Fortnine which explains the dangers of relying too much on Autnomous (?) tech in a Tesla.

This is scary to think of how careless one can be. There was another video that I had come across of some young guys playing cards in an XUV 700 while driving it. Not posting here as I don't want them to have more attention for that stupidity as much as they already have.

https://www.youtube.com/watch?v=yRdzIs4FJJg

[dust-n-bones]

This is the classic 80-20 way of tackling problems - autonomous technologies are trying to solve the 80% of driving situations today, leaving the 20% 'edge cases' to be solved incrementally.

The problem is, those edge cases are dangerous for a 'fully automated' mode in a consumer car. That however doesn't mean the 80% automation is not good enough for highway trucks (as an example) in developed countries. The opportunity to make trucks run 24*7 between hubs is probably the one of the fastest routes to commercialization of the technology.

Autonomous taxis in well known, city areas is another such opportunity, where the speeds are low enough for edge cases to be tackled by just coming to a safe stop.


So, there should be many areas where the 80% automation should be sufficient to get started and justify the enormous investments. The edge cases will be addressed slowly - based on what makes economic sense to solve first.

Automation is not 'Impossible' as the headline says, but about the art of the possible.

So will be the case of deployment this technology in under-developed infrastructures of countries like India. Unless there is an economic incentive (e.g., increased efficiency that logistics companies will value), we should not expect 'localization' anytime soon.
In our context, improving road infrastructure (quality of surface, safety, monitoring etc) will provide much higher efficiency than automation I suspect.

[Hayek]

I suspect that life saving autonomous tech (namely autonomous tech that gets into fewer accidents than the median human driver) is highly likely to be developable. But there will be a trade off between the number of accidents and the willingness of passengers to deal with excessively conservative driving, and that combined with any deaths at all arising from improperly functioning cars is something people are very unlikely to accept. It’s the liability that will accrue to producers of an autonomous car than saves even 50% of the lives lost in accidents that makes this tech impossible to implement. There were ~43,000 road deaths in the USA last year - any product that kills 20,000 people a year in America (without human intervention) will put the makers out of business very quickly

[Roadster101]

My two cents on this -

Fully autonomous cars are not made for a country like India. While driving, you literally spot a moron at the wheel every 100 meters and with people making sudden maneuvers, veering off their lanes and overtaking from the wrong side, it's very hard to program a car that can suit the driving conditions of our country. Not only the drivers but also the roads are messed up. Poor lane markings, potholes and design flaws are just a few of the many flaws flaunted by our roads.

In order for fully autonomous cars to function at its best, each and every vehicle present on the road has to be fully autonomous and that is a very costly affair and could take many years to get implemented.

P.S. Human input is always required, softwares just make our lives easier but in the end, we have to rely on ourselves cause our understanding of our roads and environment is much better judged by us rather than a software.

Cheers.

[One]

Originally Posted by GForceEnjoyer

Unless you equip cars with supercomputers running code that can fully replicate the human thought process in such split-second decision making, there will never be "fully-autonomous" cars.

Few technical aspects:
- The frame of reference is extremely small in a car driving scenario. So the decision making framework is extremely constricted.
- Within this framework, there is no need to replicate human thought while driving. We can frame the rules, implicit and explicit in human thought while driving but more importantly, augment it with new set of rules which we may want to impose.
- The result of these decision making frameworks need to work in cars not the way these frameworks get built. None of this requires processing super computers on the car end.

Once vision related issues around depth and differentiation are solved off car, split second decision making won't be the constraint.

[GForceEnjoyer]

Originally Posted by One

Once vision related issues around depth and differentiation are solved off car, split second decision making won't be the constraint.

Thanks for the clear explanation, I stand corrected.

[Kosfactor]

I really like the direction OEMs are taking with ADAS - driver assist features, however looking at Autonomous driving as the answer to poorly trained \ licensed drivers is just a bad idea.

[Tanmay007]

Oh it is only a matter of when and not if. Fully autonomous cars will absolutely happen. Getting into taxis without any drivers will happen. It might take longer for India, but it is inevitable.

The only human intervention I see in future would be in extremely rare cases when the car will request remote support automatically. Even in those cases, the car will not require any of it’s occupants to drive manually.

Sharing a couple of videos of where we are at right now. These would be enough to blow your minds.

https://youtu.be/6v036bBD31o

https://youtu.be/N4X4GMFmTb0

[vinya_jag]

There is autopilot working well in other modes of transports, such as Airplanes, Rail, Rockets. However, for these, there is a huge alternate assitance system that is working in the background.
For Air, there is an army of base stations that are working to keep these airplanes on schedule, on track, updating them about weather, etc. And are manned by two pilots. Same with Railways, there is a system in place, control stations, track changers, time keepers, Flag men, and yet there are Loco Pilots.
Also, if you noticed, the above are Track based. Other users on the same track are similar things, like a train travels on the same track as the other train, and and airplane travels on the same track as the other one that left before you, a few hours back.

For a Car trying autopilot, there are way too many unknowns. Roads in California aren't probably designed the same way as the roads in another state. Bridges are probably painted different. Some roads are 2 laned, some are 4. Some roads could be curved in such a way that oncoming traffic might seem coming on your side, and the lights could blind the camera. The car in front could be driving carelessly, or could be a learner. It could be a bike, it could be a cyclist. The vehicle in front could have had a malfunctioning lights. The car in front possibly is from another manufacturer who wrote their autopilot to perform differently.

And if your manufacturer feels confident that they have fed the AI, enough data to cover every situation in the world there is, what is it for? To do a long drive in your car without driving? I would take the train instead.

I wonder what the top tech companies in the US are after. Choosing the wrong problems to solve, one after the other.

[Jeroen]

I have a very different take on this topic, certainly the real elephant in the room is not being called out. In fact, there are two HUGE elephants in this autonomous car room.

I have no doubt that autonomous cars are on their way, so maybe it is going to take a little longer than originally anticipated. That tends to be the case with just about all new complex technology.

The main problem I often see in articles such as these, is that it is based on what human might do, e.g. anticipate for a kid running behind a ball into the road, and what humans are most likely to do / to behave.

Fact remains that humans are not particular good at driving cars to start with. You only have to look at the number of accidents, deaths and so on. You can take any country. Even in a small country such a my home country, the Netherlands, with excellent infrastructure, mandatory driving lessons prior to taking a driving exam, well maintained cars, well enforced traffic rules etc. we still have some 600-700 deaths per year in our tiny country. (population 17 million). And thousands of thousands major accidents with people ending up in hospitals and maimed for life.

The vast majority of these deaths is due to driver error one way or the other. Mechanical failure is almost unheard of these days.

So what are some of the problems with human drivers when it comes to their (driving) behaviour:
- They might drink alcohol prior to driving (some whilst driving)
- They might be on (medical) drugs
- they might talk or text on a mobile phone
- they might have had a rough night (e.g. parent with a young baby crying throughout the night), just being tired
- they might be having an argument with their fellow passenger
- they might be eating and or drinking
- they might get blinded by low sun
- they overestimate their driving ability
- they speed
- they brake too late
- they might fall asleep behind the wheel
- their attention will wander on long boring stretches of road
- etc etc

The list of items affecting human driving capabilities and goes on and on and on. It is just human nature. In many industries this was one of the reasons automation took over. (Process industry, aviation, trains, etc etc)

Short version; Human driving performance varies wildly from one individual to the next and is affected by a huge number of factors.

So yes, humans can be pretty good at handling unexpected situation. BUT, that is certainly not a given for all drivers. Take that example of the little kid running after the ball into the road. Some drivers will, correctly, assume when they see the ball coming into the road, a kid might follow. So they slow down, in a controlled fashion. Many drivers will not slow down. Some might swerve or slam on the brakes endangering other cars/drivers.

All these “exceptions” where humans are doing better than a computer are exceptions. It is far from a given that all humans will react well to such exceptions.

Computers/automation will give a much more consistent performance than humans in just about any task you can think of. In fact they tend to be much, much better at it than humans. E.g. the autopilot on my little plane, flies my aircraft in a much more precise way I can ever hope to achieve.

What is surprising is that these sort of papers don’t compare how automation does compared to humans overall. It only looks at possible exception scenario’s where automation might do less. It overlooks the fact that probably during 99,99% of the time the automation is much, much better than any human will ever be.

So here we have Elephant number 1:

Automation is never going to be perfect. But at what level of safety will we prefer automation over humans? What I see in these debates is that we expect automation to be perfect and outperform humans in every situation and under all circumstances. We certainly don’t expect that from human drivers, so why do we expect it from automation?

If automation would reduce overall fatalities by 50% I would think that would be a tremendous achievement. It still means there will be accidents, there might even be a rise in certain type of accidents, but it is a vast improvement!

And here is Elephant number 2:

I learned about this some years ago when I was spending some time at Stanford University on a management course. As part of the course we also visited and had sessions with some of the different faculties. We spend some time with the guys and girls working on autonomous cars. This sort of pretty advanced R&D Stanford work won’t make its way into commercial products for 10-12 years, so they are very much ahead of the game, or rather the current state of art of technology.

When asked what their biggest challenge was, they told us: human interaction with the automation. The notion that humans are any good at overseeing automation is beyond ridiculous. It has been proven again and again. (often at huge cost and lives) that humans suck at it.

Humans have a very short attention span. You can only spend so much time looking at a dashboard/road etc. If a machine is doing the actual driving, within minutes your mind will wander. When the machine/automation needs your attention it will take several seconds before you have assessed the situation and come up with a reasonable solution.

They actually had made a mathematical model showing, most likely, a semi-autonomous car needing oversight and occasional human assistance is more likely to meet with a crash than regular car. Again, not because the automation isn’t good enough, but due to human nature and how we behave.

If you fly at cruise altitude on autopilot, pilots mind will start to wander. Pilots will exhibit the exact same behaviour as drivers of cars. So some safety measures are included. During busy flight segments such as take off, landing, descent the auto pilot might be switched on, but the pilots will be very active and involved in what they are doing, not a problem to keep that going for a little while.

In my mind, the worst scenario is relying on humans to step in when automation fails. Because when you are driving a car and something happens you need to respond pretty quickly. One of the reasons we have so many accidents is also because, as it is, humans aren’t to good at responding quickly to unexpected situation. Some people are, most people are not and need a bit of time. But driving at only 50 km/h means you are travelling at almost 14 m/s.

If you were driving in your semi-autonomous car, looking down on your iPad and the computer signals it needs your assistance it will take several seconds before you looked up, looked around you, got your hands on the wheel, feet on the pedal and start responding. You might have driven 30-60 meters already!

Normal driving routines, keep your mind occupied and your attention focussed and routine/experience helps as well. But put your mind to a very different task then driving a car and switching back to driving the car is going to take time. Much more than normal driver reaction times.

Again, the problem with these two elephants is humans. This article is also a prime example on human behaviour, it focusses on the exception rather than the overall result. That doesn’t mean you should not be looking at these exceptions, but you need to do so in much broader context. (e.g. impact on overall accidents)

Humans are not particular good at making decisions based on scientific data. So even though there might be more than ample evidence of automation outperforming humans and saving lives, many people still prefer humans to be in control.

Of course, there is also the matter of how this autonomous cars need to be fitted into the various countries legal system, insurance etc. That is up to politician, they are humans, behave as humans (well, at least in theory) and listen to their human voters.

Jeroen

[Durango Dude]

Originally Posted by Tanmay007

Oh it is only a matter of when and not if. Fully autonomous cars will absolutely happen. Getting into taxis without any drivers will happen. It might take longer for India, but it is inevitable.

Thanks for sharing: the clips were so good to see and however tech has gone to make 'autonomous cars' a reality. Even when the autonomous interacts with humans it's so smooth plus add to that machine learning will be perfectly honed in a few years time. I've been to Shenzhen as a tourist and it has changed wonderfully and the clip and robotic voice was so nice and refined.