Tesla Model 3 Fully Autonomous ?

[sandpiper]

So, in the quarter ended almost 4 months ago, Google's cars drove on average 5,318 miles between incidents where the autonomous system had to be overridden by the operator.

Please elaborate on how this is "not even close" to perfecting self-driving technology.

Well... if you're talking about eliminating the steering wheel & pedals, I'd say this is still a long way from perfecting it. What were the potential consequences when the driver had to override? Probably mostly minor accidents or close calls? Probably some majors? Maybe a fatality?

Right now, in the US, people drive 3.2 trillion miles per year. At one "incident" per 5000 miles that's 640 million incidents per year. What is an acceptable rate? 1 million? 10 thousand? Any engineer or programmer will tell you that the cost & difficulty of eliminating bugs/issues increases exponentially as the number of bugs in the system decreases.

It's easy to say that the system has to be "as good or better than a person". But public opinion & courts are probably more sympathetic to a human who causes an accident and kills or injures themselves or others, than they would be toward a machine built by a big company. A jury intuitively understands the argument "We're all human and make errors." . They won't be so kind to a machine when a some 6 year old kid is killed on the road by an automated car. Would you like to the be the defense attorney arguing that "this was a corner case that the machine wasn't programmed for" and that, "yes, a person would likely not have run the kid over but that statistically 2 other kids were saved when this one died". It would ring a little hollow while the mother is sitting there in tears.

I'm sure that each individual google tester is really impressed with their machine as they happily tool around in California for 5000 miles. But the law of large numbers is still happily waiting on the bench and will have to be dealt with before the end of the game.

 

[Luke42]

Does no one here actually like to drive?

Depends on the situation. Driving is a chore when your trip is a chore.

I like driving on twisties. I don't like being stuck in traffic. Long stretches (8-12 hours) of prarie can get to be tedious and fatiguing, even if I kinda like the meditation of it.

 

[Booga]

Booga,
I don't believe that the implementation of self driving will require "any" additional elements added to the road (i.e. embedded sensors, etc.), the cost would be too high. Everything needs to be in the car, since there would be no way to change any significant portion of the roads to comply with what "the car" needs.

I used to wonder about how cars would handle things like wet pavement, where you literally cannot see the lane markings, or snow covered roads where everything is under a couple inches of snow. People even have trouble driving in those conditions. I read recently (forgot where) exactly how they are able to get the cars to work in those conditions. It's not GPS, since that can be off by many meters. The way they do it is that the cars that drive around before hand take high resolution LIDAR of the surroundings, and also high definition camera images of the surroundings. A combination of the two is used. If there are no lane markings visible, the software compares the current camera video feed to prior recorded video to compute distances to a set of known objects in both images. Things like street signs, benches, trees, even surrounding buildings. So even if there are no road markings visible, this level of processing can pinpoint a vehicles location to within something like 5 cm.

So if you know where the car is within that error margin, and you know where the road and all road markings are within that error margin or better, then any road markings are nothing more than a "nice to have".

I'm not saying this solves all problems, just illustrating how road markings are not as important as they might first seem.

And having said all that, I do believe that the rollout will be incremental. Perhaps happening on selected freeway lanes (carpool lanes), then migrating over to the other freeway lanes. Then making their way to non-freeway roads as the software improves. I could also envision limited geographic "autonomous regions" where cars are allowed to operate. Think "last mile to the train station" kinds of vehicles, where the roads they use to get to the destination are mapped out well in advance. I know for a fact that Uber is heavily involved in just this.

RT

That's creative - I wasn't aware of taking high resolution images/mappings of the area for the cars to consider. The challenge with a system like this, which is easily resolved IMO, is that it might require relatively frequent scans/updates to make sure the data hasn't expired. This might be due to an accident, construction, or even something like a burst fire hydrant.

I'm curious to see what the industry comes up with. I think that the industry will have to work together at some point, but it's clear that at least right now, Tesla is the leader in this.

 

[Luke42]

...Indeed, those sorts of active accident avoidance systems will be standard on most cars long before true autonomous behaviors are, and as a result the vast majority of safety gains will actually accrue long before the advent of L4 autonomy.

I test drove a 2016 Honda Civic Limited the other night that had automatic braking, adaptive cruise control, and lane keeping assist -- all on a high volume / low cost ($27K MSRP) car.

We're looking at the Civic as a commuter for my wife between now and Model 3 Delivery Day (outside of Chicago). We were impressed with the tech (even though it made me want a fully autonomous car), and it looks like a nice fit for her highway-heavy commute.

While these features aren't standard on the Civic, it is available on several different trim levels (EX, EX-T, EX-L, and Limited). I bet it will be standard on the Civic EX (and above) in a few years.

 

[brianman]

How does the AP decide course of action given a "no win" situation?

Something/someone is going to get hit. Take the hit head-on since the car knows it can survive since the crash is determined to be within it's safety parameters? Hard veer and potentially collide with a non-involved car? Hard brake and potentially be rear-ended (not due to someone following too close but because the reaction by AP is so quick, the human driver behind you cannot follow suit?)

How will it make these decisions? Dog crosses the road...wreck the car or hit the dog?

What do you do as a human?

 

[RubberToe]

What do you do as a human?

Briaman,
You are assuming that Seir is a human. How do you know that it isn't the upcoming Model 3 AI simply mulling over the tough decisions that will be required when it is turned loose in the world?

RT

 

[JeffK]

Personally, as long as it can handle my daily commute with stop and go traffic on the highway and some poorly paved curvy backroads plus roundabouts then I'm happy. I'd also prefer not to drive off a road and hit a tree or a deer even if weather conditions are horrible if at all possible.

 

[Thalass]

The driver will always be responsible...period. At least in our lifetimes. Look at the aeronautic industry. When I was flying regularly, once I was in the air the autopilot was able to fly to my destination including all course deviations, fly the approach, and I did not have to touch the controls until the flare for landing. However, if anything ever went wrong I was the one responsible...not the airplane. The technology for self landing airplanes has been around for a long time but the responsibility and liability is always on the pilot in command. I see autopilot autos in the same light.

Dan

Indeed. You may not be piloting the aircraft directly, but you are still in command of it. Airline pilots nowdays are more managers than pilots. Kinda makes the title "Captain" more appropriate, really. Captain tells computer what to do, computer says "aye aye" and does it.

Of course airline pilots are there mostly just in case something goes wrong. And what's happening lately is that the pilots haven't got as much experience because they mostly just manage the autopilot. There are few Captain Sully's left. That might be something we face on the road. Someone's car may still have a steering wheel and pedals, but would they know what to actually do if they had to take over? Even if trained, how recently have they actually driven a car?

 

[zenmaster]

Not understanding various guidance gestures/whistles from emergency workers, road construction workers, or police, the autopilot car would just sit paralyzed at intersections, crash sites and orange-cone mazes blocking all traffic. The car would not recognize various types of road debris in its path, choosing to stop instead of moving through those that are actually safe (e.g. styrofoam) or that can be easily moved aside by a person. Many types of ad hoc detour signage would be missed by the car.

Of course some of the car's limitations can be accommodated if consumers demand changes to law and convention - i.e. traffic cop must use a device to signal specific car to stop, go and to take appropriate path.

 

[JeffK]

Zenmaster, I've actually seen multiple autonomous driving demos that have addressed a few of your concerns already by mobileye and others.

 

[zenmaster]

Zenmaster, I've actually seen multiple autonomous driving demos that have addressed a few of your concerns already by mobileye and others.

For example?

 

[ttupper92618]

I really see two kinds of autonomous technology being discussed here. Some people are talking about autonomous technology as an adjunct to human operators. In this role, the technology would supplement human skill for safety purposes, and could completely stand in for the human operator for mundane driving tasks.

The other application being discussed is where the human is removed completely from the equation and where they are not even allowed to 'drive' the vehicle in any sense of the word. Proponents of this scenario imagine a world where it is either impossible to operate such a vehicle because it lacks physical controls, or because it is actually illegal to do so.

I think the former application is already arriving in the form of systems such as Tesla's autopilot feature, and I think it is obvious at this point that this technology will continue to evolve. On the other hand, I consider it very unlikely that we will see widely adopted automation of the second type at any point in the next 20 to 25 years if ever. There are many reasons for this, and I can only name a few here. That said, one must realize that it's going to take that much time at least for non-autonomous vehicles to largely be off the road merely from a replacement cycle perspective. More importantly, there are large technical problems remaining to be solved in order to allow these vehicles to be truly capable under the full range of conditions that would be required for that kind of automation. These technical challenges go way beyond the often cited concerns related to poor road conditions, the absence of physical markings, etc. and instead exist in the realm of contextual understanding and the capacity to exercise judgment.

 

[zenmaster]

I consider it very unlikely that we will see widely adopted automation of the second type at any point in the next 20 to 25 years if ever.

Agreed. It's all hype. Musk has said he wants the car to deliver itself, unassisted, to your door, cross country. I can't see how a knowledgeable engineer can make such a statement due to the humongous gap existing in real-world awareness A.I. Tech would need several successive breakthroughs - a jump many orders of magnitude ahead of today's conventions - while that progress is typically only on the order of one breakthrough every 10 years. As simple as the driving task may seem, the actual identification, judgement, and experience required to effectively predict, to plan and to negotiate is on the same order of complexity as that which would allow effective replacement of human professions.

 

[JeffK]

Cross country is extremely simple in the grand scheme of things, more so than you might think. The vast majority is highway driving which is fairly straightforward with two year old technology. The newer systems are geared more toward city driving, following signs, street markings, and street signals. Even if you left it at that alone it makes a perfect taxi like service you barely need complex parking skills. Also don't forget companies like Mobileye are collecting data from all the manufactures using their products. That's tens of millions of miles of data. The city driving tech is already being proven out today and in two years time it'll be even better. Deep neural networks can be trained to be better and faster than humans at recognition. With the sensors they can see more and react faster than any person. Youtube has a few videos on the current state of technology from mobileye and competitors.

It's not as far away as you think. Heck simply staying in your lane via a robot using neural networks was demonstrated nearly 25 years ago. We've come leaps and bounds since then.

 

[zenmaster]

The vast majority is highway driving which is fairly straightforward with two year old technology.

The actual problem of autonomous driving requirements *has never been* that which is needed to address the vast majority of driving (time/distance wise). The vast majority of driving problem solving is, comparatively speaking, infinitely more trivial to address than the much broader scope of possible situations that can occur.

 

[ttupper92618]

The 10 percent of real world driving conditions that require independent thought and judgement represent a hurdle that may easily consume 90% of the effort. Keep in mind I am not saying there will be no fully autonomous use cases. Indeed, I argue that over time as the technologies evolve these scopes will grow wider and wider. However, the realization of a go anywhere fully autonomous vehicle, as some here have suggested, is orders of magnitude greater in difficulty than assistive technologies that can offload the majority of driving tasks under limited circumstances. The idea that one can draw an arrow straight between these two problem states is naive and doesn't reflect the difficulty of these problem states.

 

[JeffK]

I recommend some of you check out the mobileye youtube channel
Mobileye

 

[Weezer Fan]

Depends on the situation. Driving is a chore when your trip is a chore.

I like driving on twisties. I don't like being stuck in traffic. Long stretches (8-12 hours) of prarie can get to be tedious and fatiguing, even if I kinda like the meditation of it.

I agree with you! Being stuck in rush hour traffic would be less irritating with autopilot. I also drive once or twice per year between Denver and Kansas City on I-70. Pure prarie the whole way! Autopilot would be fantastic for that trip.

 

[Jersey Shore Tom]

[QUO="ttupper92618, post: 1511554, member: 46381"]The 10 percent of real world driving conditions that require independent thought and judgement represent a hurdle that may easily consume 90% of the effort.[/QUOTE]
How do you explain Google's most numbers that it they are averaging over 5,000 between human interventions and only one accident ever where their car was at fault. Sure doesn't sound like 10% of real world driving can't be done done by computer yet, more like ~.5%. With the amount of data Tesla can collect everyday, if Tesla and Google shared data today it would be <.01% within a couple of months.

This is the information age, these types of systems are can and do so exponential improvement.

 

[ttupper92618]

How do you explain Google's most numbers that it they are averaging over 5,000 between human interventions and only one accident ever where their car was at fault. Sure doesn't sound like 10% of real world driving can't be done done by computer yet, more like ~.5%. With the amount of data Tesla can collect everyday, if Tesla and Google shared data today it would be <.01% within a couple of months.

This is the information age, these types of systems are can and do so exponential improvement.

It's very easy to explain that. They are:

- not driving in heavy rain
- not driving in snow
- not driving on ice
- not driving in most other kinds of inclement weather
- not driving on unimproved (dirt) county roads
- not driving in the idiosyncratic, dangerous, or life threatening conditions where complex situational awareness is required.

In short, they aren't driving in the real world conditions truly autonomous vehicles (vehicles with no human control inputs) will need to master to be useful in huge parts of this country, much less the world.

When these systems finally do master those first 5 tasks above - and I do believe they someday will - there still remains those idiosyncratic circumstances of point 6 where real judgment is required. And a truly autonomous vehicle - one with no human inputs - would need to be able to deal with them also. I'll give you some examples when I'm not tapping this out on a tiny touch keyboard.