Is a self-driving car smarter than a seven-month-old?(economist.com) |
Is a self-driving car smarter than a seven-month-old?(economist.com) |
Going back to self-driving, the main challenge on the roads seems to be the fact that anything or anyone can suddenly appear on the road in front of the car. It can be a drunk person, a slow animal (or a fast one), it can be a huge but empty cardboard box, or it can be a fridge in cardboard packaging left on the road for whatever stupid reason. The possibilities are almost literally infinite. A good FSD system should be able to assess, try to make a good prediction of the behavior (it's kind of OK to hit an empty box if I don't want to cause much discomfort to my passengers, but not OK to hit a fridge).
Hence in my opinion ML-based FSD is a dead end, always has been from the beginning. If you asked me 10 years ago I'd have told you the whole effort and billions of investments are going to get us some improved hardware at best but never a true universal self-driving system. The self-confidence of Google's executives, Tesla's and others' who repeatedly made predictions about this technology in the past decade is just astonishing. I've been thinking to myself all this time: how can they not see it? Really, where is this enthusiasm around ML coming from?
For one, those are all situations people get wrong all of the time. FSD doesn't have to be perfect, it just has to be better than humans.
Two, once FSD is better than humans in most conditions, we can build infrastructure around it. Things like FSD only lanes to reduce erratic human driving etc.
The way I see it - solving FSD on highways seems like a very solvable and economically valuable thing to solve, even in limited ways (i.e. dedicated lanes).
Time will tell.
To be fair, Google's (Waymo's) FSD is most definitely not fully ML-based, and afaik neither is Tesla's. Sure some people may speak like that's the case, but the technical approaches these companies take are not so naive.
How does a human driver determine if the box they are looking at is empty or has a fridge in it? How many times has there been no way to get to your destination without ramming refrigerator boxes? How often do you even find yourself pondering whether you should ram a box?
A good FSD system need only distinguish between obstacles to the degree that it can predict if the obstacle will remove itself or if it will need to find an alternate route, with the latter being a safe default assumption. Claiming an FSD system needs to do something even a human can't do is just moving the goal posts.
Because the ML based approaches to exactly the kind of problem you are describing are the most effective which have appeared for a long time. This isn't just hype (though the hype will paper over the still huge difficulties involved in applying this approach), it's genuinely the only approach which has much hope of working (It's also worth pointing out that these systems are almost universally a collection of subsystems each of which employs neural nets or other ML approaches to varying degrees, not a single black box).
Same as most other "enthusiasm" in this industry: HYPE!
Everything from self-driving to machine learning, TypeScript, React, Containers and everything in between is driven by hype, that is usually supported by huge companies investing in marketing to create said hype. Once the ball is rolling, it'll keep on rolling until people realize they got got, then we all move on to something else to hype. Rinse and repeat.
So, an area fully deserving interest.
Of course the enthusiasm brings a lot of unduly expectation, unfruitful trial and errors etc. - just normal expectable accompaniments. They are not the most important phenomenon, unless the interest is posed for the wrong reasons (industrial implementations pretending the technology is a universal solution).
By the way: "hype cycles" have occurred many times in the history of AI. All experience for collection.
Pioneering, squeezing the potentials out of a technology which has immensities to offer - just not necessarily ready for specific industrial applications. That, the article shows pretty well: a problem - "self driving vehicles" - was proposed and the tehcnology is imperfect for that one specifically.
The thing that self driving systems can do, lane keeping and stop and go traffic in a straight line could be done by a human Hild within a few hours of training, less if it's an auto gear box.
The fsd software is plain infantile compared to a human driver, it might keep a lane better than a human who does not have any business driving a car in the first place(blind, on drugs, sleepy , alcohol).
In the beginning, cars didn't have good brakes or suspensions and that led to accidents as well, but once cars where road safe, it didn't take humans long to accumulate the driving skills. In Europe, people get a driver's licence after 12-30 observed hours of driving practice and a theory test and are deemed road worthy, it works more or less. Fsd tech is now 10years old with God knows how many dev hours and collected data behind it and no hope for full fsd on the horizon anytime soon.
It's quite simple, if machines were better, they would have replaced humans, but the human brain and body is quite far more sophisticated than any of these fsd systems. Sure, for some tasks, machines are better suited and the change has happened.
Call me authoritarian or whatever, I would prohibit the sale of these systems for the time being. If the companies want to raise money for it, do it, don't charge the buyers upfront. And most importantly, don't release your beta ware to the streets where my kids are walking or are in another car. I didn't sign up for this last time I checked. The regulatory bodies are suspiciously quiet on the matter(slow is normal). A whitelist approach should be deployed, hell, in most of Europe, if you change as much as an exhaust system or use tinted windows, or different wheels, it's illegal unless there are homologation papers.
They are trying to copy the way humans drive, using vision and pattern recognition, but without real, high level intelligence behind it. The problem described in the article is just one aspect of it: a human will know that a bicycle is hidden behind the van, and he know how bicycles tend to behave and be ready when it pops out. An AI, most likely not, it means the reasoning of "if humans don't need X to drive, then neither does a self-riving car" is flawed, X can be lidar, annotated maps, machine-to-machine communication, etc....
I personally think that self-driving cars need every help they can get in order to make up for their lack of intelligence. They already have 100% awareness and superhuman reflexes, but if they can get things like extra sensors, that's even better. It may not get the bicycle behind the van, but it can somewhat compensate by catching it a fraction of a second earlier as it pops out, or maybe by knowing though prerecorded maps that it is a place where bikes are expected.
«How to give AI at least some semblance of that understanding - the reasoning ability of a seven-month-old child, perhaps - is now a matter of active research».
Err... «Now»?! No, that has been an explicit topic. Expert systems come to mind as one of the most explicit efforts to tackle it. And they are mentioned, at 60% - as a reference that the layman cannot decode! And an example of how the article makes academic matters seem like strategies promoted by industrial agents.
The Economist has better material. I now also have the doubt if on average it does stand without crutches - without the reader having to put it back into an historic framework of reality -, though.
Is an orange as smart as a screwdriver? Is an ant as smart as a waterfall? Nonsensical questions. There is no such thing as a universal quality of intelligence.
I cannot access the full article, but the leading paragraph says:
> BY THE AGE of seven months, most children have learned that objects still exist even when they are out of sight. Put a toy under a blanket and a child that old will know it is still there, and that he can reach underneath the blanket to get it back. This understanding, of “object permanence”, is a normal developmental milestone, as well as a basic tenet of reality.
So one could imagine the article is about "object permanence", something that is easier to compare between babies and self-driving cars. But not sure how interesting articles you can write about it (or interesting knee-jerk HN comments about said articles).
(You missed the link to the copy at archive.is :) )
I encourage you to not assume you know what others are thinking.
Where is this assertion coming from? Self driving requires planning which requires forward prediction and continuous object tracking, I highly doubt this is at all true.
"Modern AI is based on the idea of machine learning."
sigh..... while it's true a lot of the most ground-breaking advancements in AI over the past decade has been due to ML, it's not like it's the only set of techniques that are worked on as part of AI. For instance, self driving is a case in which many techniques that are used (planning, sensor fusion, filtering, etc.) are not ML-based.
There is more to self-driving than "Object Permanence", it's a bit of a false equivalency. My 5 year old can recognize letters / ABCs better than car... still wouldn't want her to drive.
Is FSD perfect? Absolutely not. I know some people who use it don't see it this way but I see it more as augmenting my driving than replacing me. For now, at least. But a 7 month old would have a negative impact on my driving.
No that's too cruel, cluster up some rats
> This understanding, of “object permanence”, is a normal developmental milestone, as well as a basic tenet of reality. It is also something that self-driving cars do not have.
That is just simply false. Nearly every team we heard technical details from has a “tracking subsystem” which integrates observations across time and sensor modalities. You cannot do that without object permanence.
How good is their object permanence? That is up for debate. Maybe there are situations particular versions from particular companies fail at. But then you should talk about these observed failures.
After all just because a healthy adult flunks a shell game we won’t conclude that they must lack object permanence.
Also how arogant it is from the writer to assume that out of the thousands and thousands of self-driving car engineers across many companies none of them thought that object permanance could be a trick worth implementing? What kind of ego one needs to write down a sweeping statement like that?
There is another similar topic which is dear to my heart. Implants that make the blind see again. I feel article about these, and self-driving car, read similarily stupid.
That’s not object permanence, that’s tracking the same object across multiple frames and tracking it as a single object. The below is part of the abstract for a paper “Learning to Track with Object Permanence” released this year that describes the difference between current tracking and the concept of object permenance:
> Tracking by detection, the dominant approach for on- line multi-object tracking, alternates between localization and re-identification steps. As a result, it strongly depends on the quality of instantaneous observations, often failing when objects are not fully visible. In contrast, tracking in humans is underlined by the notion of object permanence: once an object is recognized, we are aware of its physical existence and can approximately localize it even under full occlusions.
Not sure if Tesla has it or not, but there is a difference between object permenance and tracking objects across frames.
Clearly every tracking has to be able to re-localise among frames, otherwise it is not tracking. If you want to make a robust tracking you aim that it won't loose track even if the object is lost or obscured for a few second. These are all tuning parameters and questions of scenario. If you have a strong track of a vehicle with lots of evidence and you have strong priors about the road layout then the vehicle can disappear behind a bus for hours and you will still maintain the information that it is there. If you have a fleeting noisy observation about one pedestrian, and you don't really have a strong motion model about them (Because oh horror, pedestrians sometimes enter unmapped buildings and don't follow strict lanes!) then you might delete their track within seconds after they disappear.
So tracking creates information about objects, and how permanent they are is a tuning parameter. Some companies under some circumstances can choose to make the objects very permanent, some different companies or the same company under different circumstances can have very fleeting objects.
Given this, how much information would you need about the state of every single self driving system to write down a sentence like this confidently: "For a self-driving car, a bicycle that is momentarily hidden by a passing van is a bicycle that has ceased to exist."
I would be cautious writing such a sweeping generalisation even about bakeries and bread making, and that's a technology which has been practiced for thousands of years.
Here is what the author of the article could have done: Pick a specific failure of a specific self driving project and say "sometimes self driving cars struggle with object permanence. " It's not like you have to go far for an example. One of the root causes of the Elaine Herzberg accident was the car's inability to match her track among observations.
I agree with that. And if all that the article would do there would be nothing wrong with that.
But it goes further and makes sweeping statements about the state of self driving cars in general.
It talks about using machine learning to classify stop signs, and then writes: "Similar techniques are used to train self-driving cars to operate in traffic."
Which might be true for some systems, but not for others.
If you read this article you would think that present day self-driving cars are just a bunch of machine learning hooked up to a steering wheel. There are systems like that, but many people question their safety. Most other systems follow a hybrid approach, where there are machine learning components surrounded by old fashioned robotics. Leveraging the strong suits of both approaches.