They describe this as a third-generation vehicle, but it still requires multiple human-driver interventions over a single journey. How many more generations to reach reliable true autonomy?
My opinion as a researcher in AI/ML but without any expertise in self-driving tech:
We are a long way from where I would be willing to trust my life to self-driving cars - as a passenger, as another driver on the streets, as a cyclist, or as a pedestrian. Much farther away than these companies press releases make it seem.
Here's why. These driving algorithms are successful in large part because of data. They train their systems, such as visual recognition (what are the objects in the world around me), on millions of miles of visual data collected on the roads, most of it in California in the sunny daytime.
This means they are very likely to perform well in the average case when everything goes according to plan. And if deployed there they might live up to the hype and save thousands of lives compared to human drivers.
But now say you're in a major city in the midwest or northeast, for instance. It may be night time. It might be raining. There might be two feet of snow on the ground, narrow lanes, road signs covered up and unreadable. There may be a pedestrian crossing in dark colors. The street lines may be faded or nonexistent. There may be a street that is marked one way on the GPS map but is currently detoured the opposite direction due to construction.
There may be a policeman directing traffic. The police might pull the car over and direct it to a parking lot. There might be a fire truck or ambulance coming at an unusual time.
A computerized system trained on data can only perform well in situations very similar to its training set. But its vision will have a hard time recognizing objects it hasn't seen before. Its language processing will not understand unusual or novel road signs. Even if it recognizes the objects around it correclty, it lacks the "true" intelligence to deal with unforeseen situations falling significantly outside its training set.
I believe that cars are quite likely to run into novel situations they haven't experienced before, and I don't trust their reactions or decisionmaking in these scenarios. So I think what we have are self-driving cars that perform very well in the common, easy case, as we have already seen in numerous press releases, but are in my opinion very unpredictable in the long, fat tail of situations.
I think you are missing the most important part here. These cars are always online and share data between them. They have a detailed map of every street and every road bump and every road pole/sign that can be used for navigation. Even if everything is in snow and the camera/lidar is frozen and can't see anything, these cars know exactly where they are and where the road is from predictive navigation based on speed, direction, road shape/bumps from previous data that was collected from 1000s of passes before that on that very same road. At first AI cars will probably avoid certain areas that have not been mapped. Each car will signal any unexpected road blocks, data will be sent realtime to a human operator who will script a walkthrough in seconds. Such as "ok, you are legitimately stuck in traffic right now, just wait" or "ok, there is a crashed car ahead of you so the right-turning lane is closed, move into the left lane and you can turn right from here as an exception". There will be humans like ATC in all cases.
Police and emergency services will just coordinate with the "ATC" to pre-script routes differently depending on the situation.
Also, communicating and negotiating with human drivers. You have to do this all the time in a city like London, mostly on two way roads where there's only room for one car, due to obstacles (roadworks) or parked cars.
In complex situations, flashing lights and honks are not enough, you need to verbally communicate with other drivers.
I would eat my hat if an AI could handle these kinds of situations.
So in short, I'll believe the hype when I see a video of a full auto drive through London at rush hour.
If they're referring to the typical classification of autonomous vehicles, Level 5 is complete autonomy. Level 3 refers to human intervention required for everything but highway driving in good conditions. Of course, they may be referring to the third generation of vehicles made by Uber, which is a relatively useless term as each manufacturer will require a different number of generations before reaching full autonomy. Not really sure as it's entirely possible this is both third generation and level 3 autonomy.
It's Uber's 3rd iteration of their prototypes. The first they tested in private for 18 months, the second had those ridiculous 22 camera 6 Lidar sensor suites and was unveiled in May, and this new configuration is brought to you by Anthony Levandowski, Kalanick's new top general, and the only real reason for the Otto acquisition. "My brother from another mother" said Travis.
The Otto deal was set around incentives, it's not actually a flat 600 million dollar deal. If Levandowsky can deliver a Robotaxi OS that carries Uber to greatness, his networth could be astronomical. He's probably the top candidate in the world right now to lead an autonomous driving project to commercialization. He's got a vicious Randian streak and he's been at the bleeding edge of driverless vehicles since the Darpa days.
Currently, self-driving AIs are kind of like chess AIs, or the Watson AI that won at Jeopardy: they have some inhuman strengths that mostly compensate for their sub-human understanding of inputs.
So, for example, chess AIs are better chess players than humans because they have longer lookahead, even though they're worse at analyzing a given board. The Watson AI didn't understand questions as well as a human, and had obvious comprehension failures such as the final jeopardy answer, but when its comprehension was good enough, it had a vast database that it could perfectly recall at very high speed, and those more than compensated for its comprehension problems.
Driving AIs are not as good at understanding what's happening around them, but they are constantly attentive in 360 degrees and have fast reactions, which (may, someday, but does not yet) compensates for their imperfect understanding of the world.
This AI is not generalizable to tons of other circumstances in which there is no obvious way to parlay the inhuman strengths of the AI into compensating for their weaknesses. As such, while there may be some other places where a driving-AI-like intelligence could be used profitably, there probably aren't many such places.
Yes, but there are lots of applications for driving capabilities, that are not getting people from a to b.
Like transporting cargo at an airport, dumpers at a pit mine, even long haul trucking (see Otto).
We are a long way from where I would be willing to trust my life to self-driving cars - as a passenger, as another driver on the streets, as a cyclist, or as a pedestrian. Much farther away than these companies press releases make it seem.
Here's why. These driving algorithms are successful in large part because of data. They train their systems, such as visual recognition (what are the objects in the world around me), on millions of miles of visual data collected on the roads, most of it in California in the sunny daytime.
This means they are very likely to perform well in the average case when everything goes according to plan. And if deployed there they might live up to the hype and save thousands of lives compared to human drivers.
But now say you're in a major city in the midwest or northeast, for instance. It may be night time. It might be raining. There might be two feet of snow on the ground, narrow lanes, road signs covered up and unreadable. There may be a pedestrian crossing in dark colors. The street lines may be faded or nonexistent. There may be a street that is marked one way on the GPS map but is currently detoured the opposite direction due to construction.
There may be a policeman directing traffic. The police might pull the car over and direct it to a parking lot. There might be a fire truck or ambulance coming at an unusual time.
A computerized system trained on data can only perform well in situations very similar to its training set. But its vision will have a hard time recognizing objects it hasn't seen before. Its language processing will not understand unusual or novel road signs. Even if it recognizes the objects around it correclty, it lacks the "true" intelligence to deal with unforeseen situations falling significantly outside its training set.
I believe that cars are quite likely to run into novel situations they haven't experienced before, and I don't trust their reactions or decisionmaking in these scenarios. So I think what we have are self-driving cars that perform very well in the common, easy case, as we have already seen in numerous press releases, but are in my opinion very unpredictable in the long, fat tail of situations.