The human has windshield wipers that removes the snow so that he can see out. What do the cameras have?
The front-facing cameras are covered by the wipers. The rest are somewhat less directly exposed to the elements. I dunno about snow, but in rain at least the cameras remain usable.
False positives are annoying. But false negatives can be deadly, and I’ve had two of them in my Eacape. Once I was about to lane change, and the blind spot warning system was silent. But I shoulder check out of habit, and when I did I saw a car sitting right in my blind spot, I would have hit him had I changed lanes.
The other time was recent. I had already learned to ignore the warning system, and was backing out of my garage pad. The system was silent, but when I looked behind me there was a delivery man standing right behind the car.
In aviation you are taught that if a gauge can’t be trusted, you should cover it with something so you aren’t even tempted to use it. I feel the same way about my Escape. A warning system that only works 90% of the time is much worse than no warning system at all.
In a self-driving car, if the sensors issue a false negative, someone gets hurt.
Here’s a BBC report by a journalist who rode around San Francisco in an autonomous vehicle.
His experience was mixed, but it seemed to basically work. To be honest, I had no idea we were at the point where paying customers can take a robo-taxi.
Bloomberg recently published an article: Even after $100 billion, self-driving cars are going nowhere. (Behind paywall.)
The article is one-sided and points out what’s missing or wrong with current autonomous driving.
Then there’s this:
No mention of the actual number of deaths, which was 42,339 in 2020. And there’s more than just deaths from the 11,320,000 car accidents in 2020. There were 4,800,000 injuries serious enough that a medical professional was consulted.
(See National Safety Council Injury Facts)
Back to the article:
But no data presented.
The rest of the article is about Anthony Levandowski, “the engineer who more or less created the model for self-driving research and was, for more than a decade, the field’s biggest star.”
Jalopnik ran an unpaywalled article the other day about this Bloomberg story. BTW, the $100 billion number comes from a McKinsey & Co study of how much investors have poured into the technology.
In a world filled with breathless reporting about “AI” “self driving” cars this article seems like a breath of fresh air.
The cars standing still in their own little traffic jam that were reported about this summer are pretty much where “self driving” is today. It indicates that the only development is that authorities in the US are increasingly lax with others’ time and safety.
I agree with the article. Self-driving is in its infancy and we’re already running into ‘wicked’ problems. As usual, people greatly underestimate how special humans are and how good they are at generalizing and making rapid decisions with incomplete data.
Lots of AI people have said that self-driving will be marginal at best until we have AIs with general intelligence - and most of our AI research is not on a path towards general intelligence. Even Elon Musk is now saying that the highest levels of self-driving may not be achievable until we have an AGI.
Driving is a complex environment. Like other complex systems, it often looks simple from the highest levels, but complexity reveals itself as you dive into the details. We’re now diving into the details, and the problem seems to be getting harder, not easier.
And that $100 billion that keeps getting quoted, specifically $106 billion from the McKinsey report, is as of October, 2020.
We worked with McKinsey on some of this, and when I would ask them what the potential payback period on all this investment was, and how, they would inevitably point to robo taxis and trucking, and throw out a year like 2030 (IIRC, by the time I retired, that had inched out to 2036). Trucking did always make sense to me, because the US has had a chronic shortage of truckers, something like 80,000, but I never really got the robo taxi benefit.
I also have never thought L5 would come to fruition, but I was by far the most pessimistic of the group involved in all this.
I’m not super-impressed with the way the report has been interpreted in the articles about it. Here’s a partial breakdown:
So, $36.8B for “ADAS components”. That is, presumably, things like LIDAR, high-resolution RADAR, steering/etc. servos, and so on. But all of these things are doing to be useful beyond self-driving. Already are useful for basic lanekeeping devices.
I’m not sure what’s getting lumped into “semiconductors”, but presumably, Intel’s purchase of Mobileye for $15B would be in there. Okay–but Mobileye actually has a reasonable existing business selling low-level ADAS systems.
In short, Jalopnik and other sites are presenting the $106B figure as if it has been a complete waste so far, when really it’s encompassing highly successful systems that are being sold now. It doesn’t look like more than a small fraction of that is exclusive to L5-type systems.
Yeah, that’s a fair callout. If I could wave my magic wand to split out the pieces, I think I would want to see how much was spent on hands-free driving vs collision avoidance/emergency braking/sideview warnings etc.
Agreed, though it’s going to be hard to disentangle. My Tesla has 8 cameras. Even ignoring the driving aspect completely, the cameras are useful as dashcams and for parking security. But they’re also useful Autopilot and FSD. If someone has a business selling cameras made for self-driving use, how much do we account to mundane uses vs. self-driving? I don’t think there’s a clear answer.
Yeah, that’s the thing. Self driving requires a whole bunch of features that improve the safety of a car. Even if it’s 50 years until we have all-purpose self driving cars, the development of them is improving cars right now.
It seems that Tesla FSD is not Level 3, even though it’s called “Full Self-Driving”. How far away is it from being fully autonomous?
Tesla has two basic ASAD products, which are related but should not be confused: Autopilot and Full Self Driving (FSD).
Autopilot is level 2. It requires keeping your hands on the wheel, not just for legal reasons, but because of its inherent limits. It’s mainly for highway use, though it can stop at red lights and does ok with just following ordinary roads. Even on the highway, it’s not good enough that you can ignore your surroundings, though in my opinion it’s a great stress-reducer and allows the driver to act more “strategically”, not worrying too much about keeping in the lane, and can pay attention to events farther out (ahead and behind).
FSD can be purchased as an option, but it is currently in beta, and is still far from a “real” release. It also requires keeping your hands on the wheel, but that is mainly for legal/safety reasons. It’s capable of making turns, including unprotected left turns, with no intervention. It is very impressive with its object recognition. That said, it makes many mistakes and it will likely be some years before it can be used hands-free.
I agree with the notion that a true level-5 self-driving requires full general artificial intelligence, but that is mainly to catch the long tail of edge cases that will inevitably pop up. I think we might be a couple of years before we get some degree of hands-free driving, perhaps geofenced to some extent. It may come first for highway driving.
If instead of the difficult driving task you consider the mundane driving task, then the comparison between robots and humans can flip around. Humans will pay extra attention during difficult task, but will zone out during boring task.
The driving task where robots are doing the best, such as freeway driving, are very boring, monotonous, and result in humans engaging in other tasks. For every instance of a Tesla hitting a parked fire truck, there are thousands of incidents of distracted human drivers rear ending stopped traffic on the freeway.
I haven’t read the article, but my generic trouble with these “billions spent and no results” articles is the idea that the money and research was wasted. Often times we’re just not there yet. If we invested no money, we’d have no progress. If we knew the outcome before investing, then of course it would be stupid to throw the money away, but we don’t know the outcome.
We know the problem is hard, but so are lots of other problems. Solving hard problems can often have great rewards, which can attract investment.
I don’t think anyone would argue that ADAS isn’t beneficial. What a lot of us are arguing is that the potential for autonomous vehicles, particularly promises of L5 in the near term, has been wildly oversold, and the technical complexities of actually delivering it equally underappreciated.
If all these companies spent more time and effort just getting traditional ADAS rear end collision avoidance into more vehicles, and making it both cheaper and standard, you could cut rear end collisions by 50%. But that isn’t as sexy to investors as hands-free driving and robo taxis.
ETA apparently there’s an agreement in place to make automatic braking standard starting September 1 of this year, so that’s a big deal.
Waymo adds Los Angeles:
That did surprise me… California doesn’t let them do ride hailing without a fallback driver at the wheel, and I assumed their next stop would be somewhere a little less regulated.
Good overview of the current self driving technology.
