I don't entirely understand why with this image though.
While just a camera, looking just at the picture of the cyclist experience some perception issues due to skewing, the camera still sees the distance between the front bumper and the bicycle, but also sees the breadth of the car.
If, in this image, we see 3 bicyclists and a car in front of us, what's the worst decision we can make? To only pass the bicyclists at that car's width? To wait until we can safely pass? To treat it like a bicycle race with a support vehicle and wait to pass until we can pass the entire group?
I’d imagine the software would predict slower speed from bicycles and react accordingly, potentially backing off more than it would a car. Which might not be a bad thing, depending on how much is prediction versus reaction in these setups.
That it’s a different color recognition says something, and it might be that. Hard to say
We are already in a failure state, the camera did detect a cyclist that is not a cyclist. At this point motion detection will do something, but since the back of the car transforms differently than a real cyclist, this something is at best right by random chance. The worst decision it can make is, that the cyclists are speeding away from the autonomous car and it can therefore accelerate.
I think the image is just meant to showcase that one sensor can be confused by edge cases. It's a good image that shows unintended interpretation of data. There are likely many other edges cases that are not as clear why a car will misinterpret them, which is why they went with this image.
"Avoid highway braking" was drilled into us as 15 year olds in drivers ed. That still is one sacred rule i swear by.
Chain reaction braking, unnecessarily has caused many accidents (witnessed by me). Driving on I-95 in Florida you have (upgrades) where bridges are were built (limiting sight). Heavy congestion causes unnecessary braking. Several rear ends were witnessed on a trip north from the keys.
Never highway brake.
Im glad you posted this because it is good to know what cars may not respond the way a human was trained.
Try braking on the Merritt parkway and see what happens
But when an accident due to highway braking occurs, the person in the following vehicle will always be considered at fault -- the person in front of you may brake hard for any reason good or bad, and if you collide into the back of that vehicle, you were following too closely, driving too fast, and/or not paying attention.
The person in front may also be partially considered at fault if they didn't have a good reason, but the person following will always be considered at fault.
My drivers ed class was taught by a lunatic (aren't they all?) that was ALL about safety. Especially defensive driving. Predicting possible outcomes, having an "out", seeing through cars, avoiding driving behind box trucks and much more
Predicting what a driver will do is a large part of defensive driving.
Who is at fault matters... but it still sucks to be involved in a collision. Not being responsible is cold comfort, especially if it results in injury to passengers in your car.
Turning off safety features sounds like a bad idea. You can turn off the fusebox or circuit breaker in your house with a penny per circuit, but should you?
Having a car be able to suddenly and unexpectedly brake hard to a full stop is not always safe. As op said s/he didnt know why
Commuter and congested driving conditions by regular drivers assume a pattern of high speed. I do not want my car to have unconditional ability to brake when a computer thinks it should. Period.
Yet... We are now in phases of non-human controlled vehicles and it is a safety requirement to know what outcomes a computer can place on a car going at high speeds.
I would like to be able to create actions that cause my safety to increase against a computer assisted car. If decals (or any actions) make me safer because it makes a computer controlled car behave a certain way, then i will do it.
(my next book: how to defensively drive against self driving cars...hmm. )
To torture my analogy, If a circuit breaker keeps breaking for unknown reasons the solution is not to use a penny (the equivalent of disabling the brakes). But rather to get professional to look at and fix the system.
If the self driving feature can be completely disabled, that would be much better than just disabling the brakes. Then drive it to someone to fix the assistive driving tech instead of trying to use jerry rigged assistive driving tech.
The 2nd airbag can be disabled in the event small children are riding in rear facing car seats. This prevents the airbag from opening and killing the child by smashing him/her into the passenger seat.
Children that small should be in the back. It is the workaround for a bad situation (perhaps you have to transport 1 more kid than you back seat can handle and you don't like this one so much). This is also well understood and studied already.
There are commonalities between this and that. But in this situation wouldn't turning off assisted driving in its entirety be better.
This is my biggest worry with the current features, actually. A lot of people seem overly concerned about distracted drivers overlooking things and getting into rear end collisions that the car overlooks.
I'm more worried about false positives causing unexpected hard braking and leading to an increase in high-speed rear-end collisions on the highways. These are really uncommon now, but could increase if false positives are common.
Even if that does happen I think we would see corresponding decrease in head-on, T-bone and other more devastating collisions.
Its like non-corrupt red light cameras, they do reduce red light running and do increase the number of accidents. They reduce fatal accidents but increase nearly harmless rear-end collisions. So property damage and loss of life are reduce, but speeders and tailergaters get the majority in my town because there are no pedestrians and no one thinks they will be the ones to cause an accident.
I wasn't trying to make a positional statement. Just noting that vehicle wraps are just one of many corner cases where multiple sensor types might be useful.
I agree with you. The one silver lining is that the autonomous car will (hopefully) have super human reflexes and will be able to brake in time for some of those...hopefully...
One of the deer encounters led to an accident (worse for the deer), another one not listed also led to an accident (Volvo skidding down a slope hit traffic divider which got launched through the top of my car, instant convertible).
Human reflexes are actually quite good when you are in your twenties or so. The good news about self driving cars is that there will be less variance in reaction speed, as well as communications capabilities warning cars behind of obstacles and danger. Right now all we have to go on in this respect is brakelights but I think it is very well possible that at some point cars will purposefully communicate in some open format (broadcast) with other vehicles on the road.
Humans all having similar reflexes can be advantageous in a line of braking cars. Following a vehicle that can stop much faster than you requires a larger following distance.
> Well, is one of many reasons why. I assume there are lots of edge cases. Floating plastic bags, for example.
A friend once almost swerved dangerously to avoid some scary-looking pipe pieces on the middle of the freeway. While contemplating making his move he saw a few more pieces fall off the back of a truck further ahead and bounce - PVC. He correctly calculated that swerving at high speed in traffic was more dangerous than just hitting a lightweight piece of PVC pipe.
Yeah, it's not hard to see that it is a car while stopped, but imagine if it is parked on the side of the road in foggy conditions. You could be thirty feet from it and not see it at all.
I've never seen that before but couldn't it have been intended as a safety measure - allowing drivers stuck behind a truck a view of the road ahead so they know when it is safe to overtake?
Yes, for a human driver. An AI driver might see the lane stripes in the video and think "There is road in front of me, I can accelerate" - injuring its passengers and endangering others.
A good example of why relying on vision alone rather than also using sensors like LIDAR is silly.
Selfdriving is already going to be a hugely valuable technology even if it requires an extra few thousand dollars per vehicle for hardware. Why get greedy?
I assume Tesla skipped LIDAR not really because of greed, but because the end consumer price of the car wouldn't have made sense. The other players aren't currently trying to sell cars with LIDAR to consumers. Is there an entity that has passed on LIDAR for different reasons?
Last I heard publicly, Waymo had gotten their LIDAR kit to under $10k. And still dropping.
You're telling me that a new non-driverless car for say, $15k would outsell a (otherwise identical) driverless car for $25k? The value added by driverless is well above the cost of LIDAR.
Achieving marketable driverlessness first is going to be way more important than cutting costs slightly. Everyone else who doesn't achieve it early is going to be hosed, and no amount of cost cutting will help. They'll just become suppliers or licensees for whoever achieves and patents it first, contractually ceding the lion's share of the profits
No..I was referring to Tesla, since they are often called out for not using LIDAR. They are, however, selling direct to consumers...where that $10k matters a lot, right now.
Waymo is likely either targeting operators (not end consumers) or at least waiting until LIDAR is much cheaper before they do.
It would not be that hard to regulate against this particular edge case.
In fact it kind of makes sense - in the world where we all have on demand self driving cars delivered via an app, there's no rationale for customization anyway.
Is this something people seriously feel government regulation of is needed?
Let people paint whatever they want on their damn cars (within reason). If it triggers a poorly handled edge case that's what insurance and/or lawsuits are for.
An accident that's the result of several consecutive "on in a $largenumber" situations stacked on top of each other isn't something society should be regulation. Regulation that junk is how you wind up with blue laws that only serve to enable society to unfairly harass people decades down the line
For the record I rather like owning my own car - I can see why a lot of people (particularly in towns/cities) could live without having their own vehicle but not everyone lives in the centre of a city.
A lot of the "no one will own cars" folks need to get out of their bubble more. The typical car is a sort of moving storage locker and has often been customized for preferences and needs like carrying various types of sports gear. (Also car seats for kids.) Plus, a lot of people just like having a car that they're comfortable in.
I have talked to many people about private cars (out measuring NO2 pollution in cities) and it seems that this is the main argument why they won't give up their car. Kids is another huge argument. Handling kid's gear with rentals or public transport sucks a lot.
My current thinking goes into the direction of small private electrical golf carts for cities since those fulfill the most important requirements given to me by those who say they'd never want to give up their car. One could even make them foldable and stack them upright when parked. Slow speeds mean that not much of a safety buffer is needed in the material.
Being able to make longer trips in the same vehicle is surprisingly low on the list. It is given mainly by people with kids who want to load the vehicle in front of their house and not have to pay for several vehicles (thus having a golf cart and a normal car does not sound appealing).
Maybe someday. The problem is that there isn't really a sharp divide between dense urban core and everything else. I think about the largest city nearest to me and traffic can get pretty high speed even nearby fairly dense areas. Traffic is often jammed up but when it isn't it doesn't just move at sedate speeds.
Even inside cities, vehicles frequently need to drive at speeds up to 70kph. So while small electric golf carts (neighborhood electric vehicles) will continue to be used as niche vehicles in limited areas they're useless to most drivers.
This may depend on region. I'm in Europe, where the city speed limit is at 50 km/h and frequently reduced to 30km/h and where you can get literally stuck between buildings in a few places with a larger car.
This future is unlikely to exist. For most people owning their car will be an advantage. If you already take public transit you might not understand this, but a lot of car owners have customized their car with other things. I'm not talking cosmetic things like paint either.
Many golfers carry their clubs. The chess guy has his chess set in the car. Shopping bags from the lunch break shopping trip - all things you live in your personal car. Sure you can carry them in, but you don't want to.
Furthermore, there's an assumption that sharing vehicles lowers costs.
Parking is one reason given. Maybe. But if you don't live and/or work in a dense part of a city, parking isn't a big issue. Feel free to argue that people might be better off if there were less parking but today's reality is that a lot of people don't have any parking issues on a day to day basis.
The bigger argument is around improved utilization. But the cost of owning a car is usually more about mileage than time and shared vehicles will actually have to travel more miles because they won't always be transporting a passenger. There are some age-related costs, especially in states where it snows. Yes, some people want the latest and greatest but then they have the option of leasing today.
> If it triggers a poorly handled edge case that's what insurance and/or lawsuits are for.
Insurance claims and lawsuits are cold comfort for those severely injured in crashes.
I'm not saying regulating the wraps is the right decision, but regulating that autonomous cars must be able to properly react to such scenarios probably is.
What about wall paintings and posters? While I wouldn't mind banning ads from public streets, you can't regulate yourself into safety, the car must be robust enough to deal even with people willing to break the law.
That's true but you can make some assumptions that greatly reduce the edge cases if you apply some basic rules. e.g. I can imagine self driving only lanes with high barriers on either side, at which 200+ mph speeds are achievable.
I agree high speeds would be achievable but some fundamental aerodynamics must be solved to achieve that. Cars can catch a bit of and gain lift they are often airfoil shaped.
Its not impossible, but the goals of an F1 car are directly at odds with a consumer sedan.
This "200+ mph self-driving cars" thing gets repeated a lot on HN but I've never seen anything to suggest that it would be remotely fuel-efficient. There is a reason why today's high-speed rail rarely exceeds 200mph / 320km/h, namely that around that value, wind resistance starts to exhibit cubic growth. Even with ultra-dense automated slipstream driving, I can't imagine this being very efficient. Some citations would be greatly appreciated.
Tyres might be an even bigger problem. See the Bugatti Veyron: "It uses special Michelin PAX run-flat tyres, designed specifically to accommodate the Veyron's top speed [258mph / 415 km/h], and cost US$25,000 per set. The tyres can be mounted on the wheels only in France, a service which costs US$70,000" from https://en.wikipedia.org/wiki/Bugatti_Veyron
So your vision of self driving cars involves dedicated roads being built for 200mph speed? So essentially you think the best use case for self driving cars is to replace rail between cities?
Who do you imagine will vote for the many billions of dollars it would cost to build these self-driving Veyron tracks for the absurdly wealthy? The people who can afford 200+ mph cars can probably afford to employ drivers anyway.
It's part of the self driving vision. Veyron's could be as cheap as other high-end cars if mass produced. The point of self driving is not really about replacing drivers but making overall transportation more efficient.
The advantage of lanes over trains is that
a) it would be an order of magnitude cheaper to convert a car lane into a self driving lane than build an entire railway line
b) self driving cars can still go on normal roads at lower speeds, so you solve the last mile problem
> Veyron's could be as cheap as other high-end cars if mass produced.
Yeah, right. Lack of scale is definitely why the Veyron was so costly. It definitely isn't the difficult engineering with a W12 engine or the $75k tires to keep it from killing the passengers in the event of a flat.
Also, most people don't drive "high end cars". I hope self driving eventually isn't something restricted to the wealthy.
> The point of self driving is not really about replacing drivers but making overall transportation more efficient.
Long haul between cities is not where cars are inefficient. Cars are inefficient in traffic.
> it would be an order of magnitude cheaper to convert a car lane into a self driving lane than build an entire railway line
Not really. You're stealing lanes from cars that can't self drive or can't drive 200mph, so you're making everything worse for everyone else if you do this. You also need to build on ramps and off ramps everywhere because your 200mph lane cannot safely use the same exits as the 70mph lane next to you.
> self driving cars can still go on normal roads at lower speeds, so you solve the last mile problem
The last mile is the real problem, though. Traveling between cities rapidly is not a real problem except for shipping.
Switzerland has a few car shuttle trains ("Autoverlad") to avoid mountain passes: https://en.wikipedia.org/wiki/Car_shuttle_train#Switzerland - no advance booking required. The Eurotunnel shuttle between France and the UK is another great example (but has inflexible bookings).
High-speed car shuttle trains might be the better solution, if you insist on using cars.
Yeah of course. I'm just trying to salvage the idea (door-to-door in a car, high speed). Seems like a train for cars would be the solution. Not that it would be particularly efficient, but not as terrible as 200mph cars.
no odder than reverse cars which fortunately only tend to show up to car shows. throw in buses with full coverage wraps that can be anything from a zoo scene to a open vista. Even decorative signage at the road side needs to accounted for. There is so much we filter out automatically that we forget just how challenging coding it can be.
and still we are mainly just concentrating on on road experiences, surely DARPA has concerns wholly dealing with off road and unmarked terrain. it would be very interesting to see how one deals with a standard off road trails
If you take collectively everything you mentioned and then think of other issues, they are no longer "edge cases", these will all have to be mitigated against.
Edge cases deal with extreme parameters (min or max), what you're describing is neither and poses a significant problem that needs to be dealt with.
The decal might be a bad idea, but having trucks display a video on their rear of what a camera sees from the front is actually incredibly useful. Who are we optimizing for now: The driver who wants to get from A to B or the autonomous vehicle?
[Disclaimer: Within a 1h radius there is not a single road with two lanes; without attentive truck drivers who signal when it is safe to pass I'd be spending at least 4h more per week on the road.]
Apart from illusions, what about discriminating between genuine obstacles and things that can be safely driven over ? For example, a barrier vs some old snow. The car's software has never experienced the properties of materials by walking, jumping on, touching etc. It can't predict how materials will behave when driven over.
And this is why you don't want to be one of the first people riding self driving cars. Especially since there isn't even a way for you to manually inform the car about things like this.
I wouldn't worry about that, millions of miles have been driven by them by now. By the time you get into the first mass produced truly self driving car you will not be able to compete with it.
yeh thinking about it, cars could be updated with sensory experience via the internet from various robots around the world - so long as we find a way to transform those experiences into the sensorium of the car. For example, if a robot somewhere had felt snow with its feet and legs, it could send its experience to cars and they could transform that experience into what it would feel like under their wheels (which are the feet of cars)
I understand the argument, since car sensors are currently less robust than human eyes, so you need to use many different kinds of them to get near-human results.
However, humans get by driving reasonable safely using just vision. I'm totally a computer vision layman, but couldn't you correctly interpret this situation with binocular vision?
A good set of stereoscopic cameras would tell the computer that those bicycles are not real, and even if they were, they are running at the speed limit so there's no need to break.
Radar would do an even better job. There's no reason why cars shouldn't have radars and drive better than humans, but yes, they should be able to solve that by vision alone.
> but yes, they should be able to solve that by vision alone.
Not really. A purely vision-based system probably needs to see this example in training labeled as "the back of a car" in order to differentiate it from reality.
A sensor-diverse system could detect the metal on the car and determine it is a car without ever having seen such an example during training.
And, this is just one edge case. The number of edge cases is infinite.
The point in the article is that a diverse sensor array is advantageous because it will cover more edge cases.
A purely vision based system will always have limitations compared to a system with more kinds of sensors.
> Not really. A purely vision-based system probably needs to see this example in training labeled as "the back of a car" in order to differentiate it from reality.
Have you take a look at depth based feature recognition? It will automatically recognize everything as a car, and will tell the error of the 2D feature extraction every time.
It's also extremely CPU intensive, so people don't apply it as often. It's a clear example of AI being hold down due to limitations in computer power.
Why would you want to guess at a 3d mapping from a 2d image when you can get 3d data directly from a sensor, like lidar?
Getting 3d from 2d images is certainly interesting and will have its applications. It's less accurate and more computationally demanding than using 3d input directly, however. Cars can and should have more than 2d input.
The point of the article is that using a diverse array of sensors in cars will help the algorithm make better predictions.
There are many edge cases out there that aren't covered,and we shouldn't expect engineers to plan for them all. What we want to do is throw a lot of raw input at an algorithm and run it through a ton of simulations. All steps of this are non-trivial, but with more data, it is easier.
You have to define "shitty" in this context. I agree that there are a lot of bad drivers out there. But even with those bad drivers, people driving while texting/intoxicated, and with older vehicles that don't have all the latest safety and assistive driving features, the fatality rate in the US is still something like 1 per 100 million miles. There are a lot of auto-related deaths in the US but there are a lot of miles driven.
I think this is only partially true. When I saw the image I was actually confused by why the outlines were wrong. In real life though I wouldn't be presented with a single image, but rather high res video.
Hardware in human eyes is actually pretty bad. Modern cameras can record and process more information in higher resolution and dynamic range.
The advantage is mainly in software. For example, I have a spot on my daily commute where I need to yield, but thick vegetation blocks the line of sight. I have to look for the signs of a moving car through the bush where there are some clearings, then wait to give the cars that I can't see the time to appear on the other side, if they exist, and proceed only then. I seriously doubt an image recognition system can do this.
Well, in this case we have a clean slate, though, so we can decide what is optimal. It may be that LIDAR is more effective than a purely binocular approach to detecting distance.
At first, I interpreted the image as a mockup of a self-driving car broadcasting warnings about upcoming hazards, so that cars to its rear could know about them before directly observing them. I wonder whether that will ever be possible across differing car brands.
I really hate to be that guy, but I think my biggest worry is how each of these car manufacturers will handle such edge cases.
Each car manufacturers autonomous driving is a black box, we don't know whether said edge case is handled or not, or what action the car will take. Maybe now's not the time but it would be nice to see a deterministic rule book that says, for this edge case, the system will react this way. We can't account for all edge cases but some default fall back would be nice.
We don't allow our plumbing, structure wiring, welding, or any other physical construction method to be closed away and hidden from view. Rather we demand these things be inspected, certified, documented and generally adhere to demonstrable codes and rules.
Keeping you source hidden is saying "I have something to hide". Why should someone have something to hide in Pacemaker software, software driving vehicles or safety system on nuclear power plants or anything else where life could be lost?
I think in some municipalities you can walk down to the planning and get most the blueprints you would like.
Where you can't it is partially because the analogy breaks down physical is slightly reduced by openness, its allows attackers of all kinds to plan better. More manpower can be focused to overcome and issue and a map is force multiplier. Security in information systems is strictly increased, because the flaws exist open or close, but when exposed in the open they are closed forever to all attacker regardless of manpower.
We also have a concept of how to inspect plumbing. I could put all rules on a checklist and teach you how to check the plumbing in your house. (I'd have to look the rules up, but conceptually I could).
Of course houses are all the same: a few toilets, sinks, tubs and showers. The differences are easy to handle in the rules. You could imagine something different: instead of water we have a special hand cleaner in this sink. My checklists cover the potable water system and the non-potable water drain system, this special cleaning fluid isn't covered and your cannot inspect it. We have a simple exception for those cases: a professional engineer can legally stamp it saying "I have thought of everything, this will work if installed according to these specs." Then the checklist is make sure that everything is up to those specs.
I honestly do not know how to create a checklist for source code. Sure I can create some starts, but https://en.wikipedia.org/wiki/Underhanded_C_Contest needs to be considered: can someone pass your checklist and still win?
Maybe we need an open source protocol that has all the standards and edge cases handled.
I think competitiveness is a good thing and will drive innovation. But in this specific field, with regards to safety and security, it needs to be out in the open.
Comma.ai wasn't originally going to be open-source, Comma.ai intended was going to be a $999 proprietary retro-fit product. NHTSA (National Highway Traffic Safety Administration) asked Hotz to delay the launch on safety grounds.
He then open-sourced the software specifically to avoid regulation and safety standards.
"NHTSA only regulates physical products that are sold,” Hotz said.
“They do not regulate open source software, which is a whole lot more like speech.”
>> But in this specific field, with regards to safety and security, it needs to be out in the open.
This is one of the horrors of capitalism.
If there's money to be made, you can bet that if someone creates a proprietary protocol, it won't be released and will force other companies to develop their own technology to compete.
Yes, it will drive competition, but I fear that it will spawn a host of different, competing technologies and make the whole industry awash in proprietary technology. In most of those cases, the people who lose are the consumers.
John Deere (I work for John Deere but do not speak for them) freely licenses all safety patents except for ones that we think any competent engineer can work around. We figure we make more money from someone who buys our competitors product and because of a safety feature survives to make the decision on what to replace it with when it wears out; than we make from the person who buys something without a safety feature and as a result dies.
I used to work for the company that published rfc 1223, giving away private information. The motivation is a competitor reversed engineered the protocol, and some customers had both. There was now a need to change the protocol but we knew our customers would not accept having to throw out the competitors equipment to take the upgrade. By publishing the protocol we ensured the competitor understood it and thus when the bit that was "always 0 apparently unused" changed to 1, with a change to the meaning of all following bytes they understood why it was always 0 and why the change happened and how to read all the following bytes.
The protocol is likely uninteresting rather it's the models, training data, and training configurations which are most interesting and the least likely to be open sourced. I don't work in the autonomous vehicle space but I would be surprised there's a set of deterministic rules that are easily reviewed instead collections of deep nn models.
The data is the real rub. There's some open source code in the space and we'll see what else gets opened up over time. But AFAIK, no one is talking about making their data sets public (and, to the degree that data is coming from regular drivers, there would probably be privacy concerns with doing so). And, as you say, with NN techniques, a lot of actions aren't deterministic or understandable.
Wouldn't you test the output rather than the input? Before an autonomous vehicle / update is allowed to go on public roads, it must pass a "driving test" that includes a bunch of known edge cases.
Sure. And maybe that's what will happen. But now you have pre-defined tricky situations. Everyone builds those very specific situations into their systems but may not really generalize them.
This happens to humans as well. Case in point: radio shows/ads that have car horn noises
The decal is bad but the car doesn't risk doing anything that it shouldn't (like driving through it). If it was a decal of an empty road ahead it would have been more problematic
Any car can break suddenly for any reason or no reason at all. It's not a tragic failure mode, instead, the desired situation is that all failures lead to this.
But that's not special. Vehicle brakes all the time and they/we don't optimize for a tailgater. That's an optimization to be deferred to the tailgater.
The decal might be deliberately constructed (using perspective tricks) to alter the behavior of an autonomous car following your van. The cyclists might appear either much closer or farther than the plane they're drawn on. If you install a video panel on the back of your van, there are even more possibilities. You might be able to play a video of an object rapidly moving towards the autonomous car, which will cause it to swerve to try to avoid it.
And yes, some of this stuff will probably work on humans as well, but autonomous driving is supposed to be better than humans.
Deliberately malicious behavior would land such person in jail as soon as serious accident happens. You can do a many things to trick human drivers as well to cause an accident, the only reason we don't see this happening very often is because most people in society are decent human beings.
> The offender could claim whatever mod he made wasn't intended to cause harm to others. Then you have to prove intent, which can be difficult.
Or you have to have laws which cover acts done with mental states short of intent to cause harm like (in ascending I order of seriousness in existing law—and all of these are already covered for acts that get people killed) criminal negligence, recklessness, and depraved indifference to human life.
Of course, it's a lot less difficult to convince even a criminal jury of intent than people in HN often seem to think; legal proof, even to meet the “beyond a reasonable doubt” standard, is far less than logical proof.
And what would be the point of a decal/video screen designed to trick people (Autonomous or not) into rear-ending your van?
That said I do agree with the premise of the article, in order to beat humans cars will need more sensors. As humans we have our two eyes behind a windshield that can be blocked by fog, heavy rain/snow, the glare from the sun, or distracted by a litany of other things.
By having multiple camera sensors in addition to things like radar, lidar and ultrasonic sensors it should be "easy" to beat humans at driving, something that the human body wasn't designed for.
> And what would be the point of a decal/video screen designed to trick people (Autonomous or not) into rear-ending your van?
There is a "popular" scam in China, where someone pretends to get hit by a car and then demands compensation. This occasionally leads to funny videos where they fall down too early and the car just drives around them; but I guess the scammers still make a profit.
Something similar could be done by luring autonomous cars into a crash and then trying to squeeze them for money.
I believe this is illegal in my country (Netherlands). Also, the emergency services here can transmit a wide band FM signal with the siren sound, so you hear it playing over your radio when they come close.
I have been deaf in one ear for a bit (a few weeks) a while ago (aftereffects of a flu), I did not realize before then how much I rely on my hearing during driving. Do self driving cars use audio sensors?
I think we rely on our hearing because we can only see in one direction. With an autonomous car, you can have sensors pointing in every direction, so visual/radar/lidar clues might be all you need.
It not only reflects very well but the way some low frequencies spread out is enough to allow certain vehicles to be heard before they are visible (solid concrete could obscure the view).
this almost throwaway line at the end caught my eye: "The safest bet, then, is for automakers to use an array of sensors, in order to build redundancy into their systems."
experienced engineers know that redundancy is a double-edged sword. what do you do when the data from redundant systems disagree? having more data sources means dealing with more edge cases. a single reliable data source (as in the limitations and biases are well known) is better than two less predictable ones.
i'm not saying that that's necessarily the case here, but it's not as obvious as this line suggests that the answer is just to add more kinds of sensors (and thereby add more diverse data sources).
I disagree. In an autonomous system, more sensors is almost always better, provided you have enough computing power to process the data. Having two sensors that disagree gives you much more information than if you had a single sensor with a malfunction. In the first case, you know there's some kind of fault and can take steps to fix it. In the second, you're running off bad data and who knows what could happen.
Three sensors are better than two—it's a common setup in flight control systems to have three identical computers running all calculations, and the majority opinion is taken as truth.
> what do you do when the data from redundant systems disagree?
Self driving car systems' rules aren't directly decided by humans. The system is trained to come up with its own rules given sensor input. So ideally, you have more different types of input from diverse sources. Machine learning is a lot easier when you have a diverse dataset.
> having more data sources means dealing with more edge cases.
Yes, and that's a good thing. You want the system to be able to handle the edge case from the article, and a diverse array of sensors can do that.
> a single reliable data source (as in the limitations and biases are well known) is better than two less predictable ones.
Can you provide an example of a single vehicle sensor that would be preferable to two others? I can't see why this decision would be made, aside from attempts to lower cost. Two sensors provides more reliability than one.
the idea that more is always better is the common misperception that i'm addressing. if you've ever tried to design a system with multiple, sometimes conflicting inputs, complexity is a real cost to think about (not to mention financial cost).
two sensors are often in a master-slave relationship to avoid conflicts, and as catbird mentioned, if you need peering redundancy, then three is much better than two (but that's not perfect either), so that there's essentially error correction built in.
machine learning is cool, but it's not a panacea. there are plenty of meticulously trained AI's that fail comically in novel situations (like the example in the original article). afaict, we still don't have a lot of knowledge around the boundary conditions of the machine learning systems themselves.
> the idea that more is always better is the common misperception that i'm addressing
It's not a misperception in the case of machine learning. The system can easily discard data that isn't useful at a given point in time. When it is determined to be useful, it's great to have it there.
I'd suggest trying out some machine learning yourself to get a good understanding of why more data is better. Tutorials for Kaggle's Titanic competition can be a good primer.
> if you've ever tried to design a system with multiple, sometimes conflicting inputs, complexity is a real cost to think about (not to mention financial cost).
Not when the system is using machine learning, particularly neural networks.
All the self driving car systems are using deep convolutional NNs, so the complexity of more data is helpful, not a hindrance.
> machine learning is cool, but it's not a panacea. there are plenty of meticulously trained AI's that fail comically in novel situations (like the example in the original article)
The failure in the article was due to lack of diverse sensor input, not because the system was built using machine learning.
You're right that machine learning isn't a panacea. The limits aren't well defined. Still, in the case of self driving cars, deep learning is state of the art. Nobody is hand-coding rules for how to drive the car. It's all about having as much good data as you can gather.
>what do you do when the data from redundant systems disagree
That's a luxury problem. It's like saying that you're better off without ECC because you don't have to deal with memory corruption. With a single sensor, you still have to deal with inaccuracies, the difference is that with only one you just don't know if you have it or not.
the luxury is adding every kind of sensor without considering cost, engineering time, computing power, etc.
complexity is extra-linear, so you have to have to do the cost-benefit analysis to determine if it's worth it. and it may well be, but my point was that it's not a knee-jerk answer to always add more sensors. there is clear benefit, but is it worth the cost?
Except in this case there's no "single reliable data source". All of the useful data sources are limited in some way. You need a diverse array of data sources to fill in the gaps of each individual source.
Not to mention that one should always strive to avoid a single point of failure. A "single reliable data source" is - whaddya know - a single point of failure. If that data source stops being "reliable", then you're screwed. When you're screwed in a giant hunk of metal rolling down a highway at 65+ miles per hour, then the only system you have left is to pray to your god(s) for a good afterlife.
Does anyone have any good research or review articles into the current state of sensor fusion?
To me as a layman, Car AI really seems like one of the most "intelligence" type of AI. Yes Alpha Go and Deep Blue are really good at something that was once thought "easy" for humans and hard for computers, but AI cars seems like trying to construct part of an animal's nervous system and sticking it into a vehicular body.
Personally I think medical diagnostics are the most interesting and useful AI.
Radiology, and healthcare overall, is set to be improved rapidly if we can get some more labeled data. The hardware and software is already ready to start learning how to diagnose cancer and other health issues based on human-scan data.
Not sure whether an "intelligence" rating is possible, given that AI systems are still designed to be domain-specific, i.e., self driving cars aren't detecting cancer.
When image processing and pattern matching driven by the need to identify pornographic images by social media companies masquerading as AI meets the real world.
This just underlines what AI actually means and what is currently trying to pass for AI.
Not sure why all the defense of "humans make this mistake too" in this thread.
Isn't the point of SDC that they are better than humans? Shouldn't we support a diverse array of sensors to cover all the eventualities that we can? If some form of radar + camera solves this problem, what's the point of arguing humans would get confused by it too?
To be clear, the article alludes to the existence of three types of cars in the wild,
(1) Cars with many types of sensors
(2) Cars with some types of sensors
(3) Cars with no sensors
All other things being equal, why are people voting for (2) rather than (1)?
How about when photons flood the camera/sensor ( the sun shines it's rays directly at you ) -- or - when photons cover the subject that the camera/sensor is supposed to be reading -- last this happened the rider died ( https://www.theguardian.com/technology/2016/jul/01/tesla-dri... ) -- they're too many natural occurring factors at play to predetermine what a moving vehicle should or shouldn't do every second - for this reason - vehicles we travel in should be put in an environment first that blocks distractions - like Musk with his Boring machine - this controlled environment makes more sense - otherwise I need an alarm to wake me up when the sensors know they can't tell what's going on - I need to regain control of my vessel and guide it correctly through the photons - yes?
"When’s a pedestrian not a pedestrian? When it’s a decal"
This sub-title is without context. Would be nice if the title here matched the article title, "This Image Is Why Self-Driving Cars Come Loaded with Many Types of Sensors"
159 comments
[ 0.23 ms ] story [ 210 ms ] threadWhile just a camera, looking just at the picture of the cyclist experience some perception issues due to skewing, the camera still sees the distance between the front bumper and the bicycle, but also sees the breadth of the car.
If, in this image, we see 3 bicyclists and a car in front of us, what's the worst decision we can make? To only pass the bicyclists at that car's width? To wait until we can safely pass? To treat it like a bicycle race with a support vehicle and wait to pass until we can pass the entire group?
That it’s a different color recognition says something, and it might be that. Hard to say
We can't know how the system will react in a failed state because engineers will not have prepared for such an eventuality (ignoring negligence).
The point is there are many such unseen sitauations that current systems don't handle, and that diversifying the sensor array can help handle those.
Chain reaction braking, unnecessarily has caused many accidents (witnessed by me). Driving on I-95 in Florida you have (upgrades) where bridges are were built (limiting sight). Heavy congestion causes unnecessary braking. Several rear ends were witnessed on a trip north from the keys.
Never highway brake.
Im glad you posted this because it is good to know what cars may not respond the way a human was trained.
Try braking on the Merritt parkway and see what happens
The person in front may also be partially considered at fault if they didn't have a good reason, but the person following will always be considered at fault.
Predicting what a driver will do is a large part of defensive driving.
Commuter and congested driving conditions by regular drivers assume a pattern of high speed. I do not want my car to have unconditional ability to brake when a computer thinks it should. Period.
Yet... We are now in phases of non-human controlled vehicles and it is a safety requirement to know what outcomes a computer can place on a car going at high speeds.
I would like to be able to create actions that cause my safety to increase against a computer assisted car. If decals (or any actions) make me safer because it makes a computer controlled car behave a certain way, then i will do it.
(my next book: how to defensively drive against self driving cars...hmm. )
If the self driving feature can be completely disabled, that would be much better than just disabling the brakes. Then drive it to someone to fix the assistive driving tech instead of trying to use jerry rigged assistive driving tech.
Children that small should be in the back. It is the workaround for a bad situation (perhaps you have to transport 1 more kid than you back seat can handle and you don't like this one so much). This is also well understood and studied already.
There are commonalities between this and that. But in this situation wouldn't turning off assisted driving in its entirety be better.
I'm more worried about false positives causing unexpected hard braking and leading to an increase in high-speed rear-end collisions on the highways. These are really uncommon now, but could increase if false positives are common.
Its like non-corrupt red light cameras, they do reduce red light running and do increase the number of accidents. They reduce fatal accidents but increase nearly harmless rear-end collisions. So property damage and loss of life are reduce, but speeders and tailergaters get the majority in my town because there are no pedestrians and no one thinks they will be the ones to cause an accident.
- flying wheelbarrow
- unaccompanied wheel
- numerous deer
- toddler on a tricycle from in between two cars
- tourist stepping backwards of a curb
- oncoming traffic on a one way highway (Romania, fairly normal apparently)
- people driving backwards on a highway
- pedestrians in a vehicle tunnel
- tourbus wheel exploding
- truck wheel delaminating
- all manner of debris, from boards to clothing and large chunks of metal
I will take your floating plastic bag over any of those. And yes, I drive too much.
Human reflexes are actually quite good when you are in your twenties or so. The good news about self driving cars is that there will be less variance in reaction speed, as well as communications capabilities warning cars behind of obstacles and danger. Right now all we have to go on in this respect is brakelights but I think it is very well possible that at some point cars will purposefully communicate in some open format (broadcast) with other vehicles on the road.
A friend once almost swerved dangerously to avoid some scary-looking pipe pieces on the middle of the freeway. While contemplating making his move he saw a few more pieces fall off the back of a truck further ahead and bounce - PVC. He correctly calculated that swerving at high speed in traffic was more dangerous than just hitting a lightweight piece of PVC pipe.
Apparently painting vehicles grey can cause people not to see them as well. Having photo realistic images of road scenes seems much worse.
Have they never watched roadrunner cartoons?
The painting lasted less than 24 hours though.
http://www.e-farsas.com/parede-desenhada-com-o-tunel-papa-le...
[1] https://i.imgur.com/AQCbuZp.gif
Selfdriving is already going to be a hugely valuable technology even if it requires an extra few thousand dollars per vehicle for hardware. Why get greedy?
I assume Tesla skipped LIDAR not really because of greed, but because the end consumer price of the car wouldn't have made sense. The other players aren't currently trying to sell cars with LIDAR to consumers. Is there an entity that has passed on LIDAR for different reasons?
You're telling me that a new non-driverless car for say, $15k would outsell a (otherwise identical) driverless car for $25k? The value added by driverless is well above the cost of LIDAR.
Achieving marketable driverlessness first is going to be way more important than cutting costs slightly. Everyone else who doesn't achieve it early is going to be hosed, and no amount of cost cutting will help. They'll just become suppliers or licensees for whoever achieves and patents it first, contractually ceding the lion's share of the profits
Waymo is likely either targeting operators (not end consumers) or at least waiting until LIDAR is much cheaper before they do.
Perhaps not a crash, but I could see AI avoiding it by hanging behind.
In fact it kind of makes sense - in the world where we all have on demand self driving cars delivered via an app, there's no rationale for customization anyway.
Is this something people seriously feel government regulation of is needed?
Let people paint whatever they want on their damn cars (within reason). If it triggers a poorly handled edge case that's what insurance and/or lawsuits are for.
An accident that's the result of several consecutive "on in a $largenumber" situations stacked on top of each other isn't something society should be regulation. Regulation that junk is how you wind up with blue laws that only serve to enable society to unfairly harass people decades down the line
My current thinking goes into the direction of small private electrical golf carts for cities since those fulfill the most important requirements given to me by those who say they'd never want to give up their car. One could even make them foldable and stack them upright when parked. Slow speeds mean that not much of a safety buffer is needed in the material.
Being able to make longer trips in the same vehicle is surprisingly low on the list. It is given mainly by people with kids who want to load the vehicle in front of their house and not have to pay for several vehicles (thus having a golf cart and a normal car does not sound appealing).
Maybe someday. The problem is that there isn't really a sharp divide between dense urban core and everything else. I think about the largest city nearest to me and traffic can get pretty high speed even nearby fairly dense areas. Traffic is often jammed up but when it isn't it doesn't just move at sedate speeds.
Many golfers carry their clubs. The chess guy has his chess set in the car. Shopping bags from the lunch break shopping trip - all things you live in your personal car. Sure you can carry them in, but you don't want to.
Parking is one reason given. Maybe. But if you don't live and/or work in a dense part of a city, parking isn't a big issue. Feel free to argue that people might be better off if there were less parking but today's reality is that a lot of people don't have any parking issues on a day to day basis.
The bigger argument is around improved utilization. But the cost of owning a car is usually more about mileage than time and shared vehicles will actually have to travel more miles because they won't always be transporting a passenger. There are some age-related costs, especially in states where it snows. Yes, some people want the latest and greatest but then they have the option of leasing today.
Insurance claims and lawsuits are cold comfort for those severely injured in crashes.
I'm not saying regulating the wraps is the right decision, but regulating that autonomous cars must be able to properly react to such scenarios probably is.
Its not impossible, but the goals of an F1 car are directly at odds with a consumer sedan.
Tyres might be an even bigger problem. See the Bugatti Veyron: "It uses special Michelin PAX run-flat tyres, designed specifically to accommodate the Veyron's top speed [258mph / 415 km/h], and cost US$25,000 per set. The tyres can be mounted on the wheels only in France, a service which costs US$70,000" from https://en.wikipedia.org/wiki/Bugatti_Veyron
Who do you imagine will vote for the many billions of dollars it would cost to build these self-driving Veyron tracks for the absurdly wealthy? The people who can afford 200+ mph cars can probably afford to employ drivers anyway.
The advantage of lanes over trains is that
a) it would be an order of magnitude cheaper to convert a car lane into a self driving lane than build an entire railway line b) self driving cars can still go on normal roads at lower speeds, so you solve the last mile problem
Yeah, right. Lack of scale is definitely why the Veyron was so costly. It definitely isn't the difficult engineering with a W12 engine or the $75k tires to keep it from killing the passengers in the event of a flat.
Also, most people don't drive "high end cars". I hope self driving eventually isn't something restricted to the wealthy.
> The point of self driving is not really about replacing drivers but making overall transportation more efficient.
Long haul between cities is not where cars are inefficient. Cars are inefficient in traffic.
> it would be an order of magnitude cheaper to convert a car lane into a self driving lane than build an entire railway line
Not really. You're stealing lanes from cars that can't self drive or can't drive 200mph, so you're making everything worse for everyone else if you do this. You also need to build on ramps and off ramps everywhere because your 200mph lane cannot safely use the same exits as the 70mph lane next to you.
> self driving cars can still go on normal roads at lower speeds, so you solve the last mile problem
The last mile is the real problem, though. Traveling between cities rapidly is not a real problem except for shipping.
High-speed car shuttle trains might be the better solution, if you insist on using cars.
and still we are mainly just concentrating on on road experiences, surely DARPA has concerns wholly dealing with off road and unmarked terrain. it would be very interesting to see how one deals with a standard off road trails
Edge cases deal with extreme parameters (min or max), what you're describing is neither and poses a significant problem that needs to be dealt with.
[Disclaimer: Within a 1h radius there is not a single road with two lanes; without attentive truck drivers who signal when it is safe to pass I'd be spending at least 4h more per week on the road.]
However, humans get by driving reasonable safely using just vision. I'm totally a computer vision layman, but couldn't you correctly interpret this situation with binocular vision?
Radar would do an even better job. There's no reason why cars shouldn't have radars and drive better than humans, but yes, they should be able to solve that by vision alone.
Not really. A purely vision-based system probably needs to see this example in training labeled as "the back of a car" in order to differentiate it from reality.
A sensor-diverse system could detect the metal on the car and determine it is a car without ever having seen such an example during training.
And, this is just one edge case. The number of edge cases is infinite.
The point in the article is that a diverse sensor array is advantageous because it will cover more edge cases.
A purely vision based system will always have limitations compared to a system with more kinds of sensors.
Have you take a look at depth based feature recognition? It will automatically recognize everything as a car, and will tell the error of the 2D feature extraction every time.
It's also extremely CPU intensive, so people don't apply it as often. It's a clear example of AI being hold down due to limitations in computer power.
Getting 3d from 2d images is certainly interesting and will have its applications. It's less accurate and more computationally demanding than using 3d input directly, however. Cars can and should have more than 2d input.
The point of the article is that using a diverse array of sensors in cars will help the algorithm make better predictions.
There are many edge cases out there that aren't covered,and we shouldn't expect engineers to plan for them all. What we want to do is throw a lot of raw input at an algorithm and run it through a ton of simulations. All steps of this are non-trivial, but with more data, it is easier.
It's news when a self-driving car runs a red light, but pretty normal for humans.
The advantage is mainly in software. For example, I have a spot on my daily commute where I need to yield, but thick vegetation blocks the line of sight. I have to look for the signs of a moving car through the bush where there are some clearings, then wait to give the cars that I can't see the time to appear on the other side, if they exist, and proceed only then. I seriously doubt an image recognition system can do this.
Each car manufacturers autonomous driving is a black box, we don't know whether said edge case is handled or not, or what action the car will take. Maybe now's not the time but it would be nice to see a deterministic rule book that says, for this edge case, the system will react this way. We can't account for all edge cases but some default fall back would be nice.
You're required to be given the chance to read contracts before you sign them, I don't see why software should be different.
We don't allow our plumbing, structure wiring, welding, or any other physical construction method to be closed away and hidden from view. Rather we demand these things be inspected, certified, documented and generally adhere to demonstrable codes and rules.
Keeping you source hidden is saying "I have something to hide". Why should someone have something to hide in Pacemaker software, software driving vehicles or safety system on nuclear power plants or anything else where life could be lost?
Where you can't it is partially because the analogy breaks down physical is slightly reduced by openness, its allows attackers of all kinds to plan better. More manpower can be focused to overcome and issue and a map is force multiplier. Security in information systems is strictly increased, because the flaws exist open or close, but when exposed in the open they are closed forever to all attacker regardless of manpower.
Of course houses are all the same: a few toilets, sinks, tubs and showers. The differences are easy to handle in the rules. You could imagine something different: instead of water we have a special hand cleaner in this sink. My checklists cover the potable water system and the non-potable water drain system, this special cleaning fluid isn't covered and your cannot inspect it. We have a simple exception for those cases: a professional engineer can legally stamp it saying "I have thought of everything, this will work if installed according to these specs." Then the checklist is make sure that everything is up to those specs.
I honestly do not know how to create a checklist for source code. Sure I can create some starts, but https://en.wikipedia.org/wiki/Underhanded_C_Contest needs to be considered: can someone pass your checklist and still win?
I think competitiveness is a good thing and will drive innovation. But in this specific field, with regards to safety and security, it needs to be out in the open.
He then open-sourced the software specifically to avoid regulation and safety standards.
[1] https://qz.com/849669/to-dodge-government-regulation-self-dr...[2] https://www.theverge.com/2016/11/30/13779336/comma-ai-autopi...
This is one of the horrors of capitalism.
If there's money to be made, you can bet that if someone creates a proprietary protocol, it won't be released and will force other companies to develop their own technology to compete.
Yes, it will drive competition, but I fear that it will spawn a host of different, competing technologies and make the whole industry awash in proprietary technology. In most of those cases, the people who lose are the consumers.
I'd be interested to hear if there are more recent examples.
I used to work for the company that published rfc 1223, giving away private information. The motivation is a competitor reversed engineered the protocol, and some customers had both. There was now a need to change the protocol but we knew our customers would not accept having to throw out the competitors equipment to take the upgrade. By publishing the protocol we ensured the competitor understood it and thus when the bit that was "always 0 apparently unused" changed to 1, with a change to the meaning of all following bytes they understood why it was always 0 and why the change happened and how to read all the following bytes.
The decal is bad but the car doesn't risk doing anything that it shouldn't (like driving through it). If it was a decal of an empty road ahead it would have been more problematic
Autonomous vehicle target being tailgated, operate device in front vehicle and autonomous vehicle slams on its brakes and gets rear-ended.
And yes, some of this stuff will probably work on humans as well, but autonomous driving is supposed to be better than humans.
Edit: In the real world, this is more likely to happen by accident than intent.
No regulations exist yet.
The offender could claim whatever mod he made wasn't intended to cause harm to others. Then you have to prove intent, which can be difficult.
Or you have to have laws which cover acts done with mental states short of intent to cause harm like (in ascending I order of seriousness in existing law—and all of these are already covered for acts that get people killed) criminal negligence, recklessness, and depraved indifference to human life.
Of course, it's a lot less difficult to convince even a criminal jury of intent than people in HN often seem to think; legal proof, even to meet the “beyond a reasonable doubt” standard, is far less than logical proof.
That said I do agree with the premise of the article, in order to beat humans cars will need more sensors. As humans we have our two eyes behind a windshield that can be blocked by fog, heavy rain/snow, the glare from the sun, or distracted by a litany of other things.
By having multiple camera sensors in addition to things like radar, lidar and ultrasonic sensors it should be "easy" to beat humans at driving, something that the human body wasn't designed for.
There is a "popular" scam in China, where someone pretends to get hit by a car and then demands compensation. This occasionally leads to funny videos where they fall down too early and the car just drives around them; but I guess the scammers still make a profit.
Something similar could be done by luring autonomous cars into a crash and then trying to squeeze them for money.
It not only reflects very well but the way some low frequencies spread out is enough to allow certain vehicles to be heard before they are visible (solid concrete could obscure the view).
experienced engineers know that redundancy is a double-edged sword. what do you do when the data from redundant systems disagree? having more data sources means dealing with more edge cases. a single reliable data source (as in the limitations and biases are well known) is better than two less predictable ones.
i'm not saying that that's necessarily the case here, but it's not as obvious as this line suggests that the answer is just to add more kinds of sensors (and thereby add more diverse data sources).
Three sensors are better than two—it's a common setup in flight control systems to have three identical computers running all calculations, and the majority opinion is taken as truth.
in your 2-sensor scenario, how do you know which one is providing incorrect data?
Self driving car systems' rules aren't directly decided by humans. The system is trained to come up with its own rules given sensor input. So ideally, you have more different types of input from diverse sources. Machine learning is a lot easier when you have a diverse dataset.
> having more data sources means dealing with more edge cases.
Yes, and that's a good thing. You want the system to be able to handle the edge case from the article, and a diverse array of sensors can do that.
> a single reliable data source (as in the limitations and biases are well known) is better than two less predictable ones.
Can you provide an example of a single vehicle sensor that would be preferable to two others? I can't see why this decision would be made, aside from attempts to lower cost. Two sensors provides more reliability than one.
two sensors are often in a master-slave relationship to avoid conflicts, and as catbird mentioned, if you need peering redundancy, then three is much better than two (but that's not perfect either), so that there's essentially error correction built in.
machine learning is cool, but it's not a panacea. there are plenty of meticulously trained AI's that fail comically in novel situations (like the example in the original article). afaict, we still don't have a lot of knowledge around the boundary conditions of the machine learning systems themselves.
It's not a misperception in the case of machine learning. The system can easily discard data that isn't useful at a given point in time. When it is determined to be useful, it's great to have it there.
I'd suggest trying out some machine learning yourself to get a good understanding of why more data is better. Tutorials for Kaggle's Titanic competition can be a good primer.
> if you've ever tried to design a system with multiple, sometimes conflicting inputs, complexity is a real cost to think about (not to mention financial cost).
Not when the system is using machine learning, particularly neural networks.
All the self driving car systems are using deep convolutional NNs, so the complexity of more data is helpful, not a hindrance.
> machine learning is cool, but it's not a panacea. there are plenty of meticulously trained AI's that fail comically in novel situations (like the example in the original article)
The failure in the article was due to lack of diverse sensor input, not because the system was built using machine learning.
You're right that machine learning isn't a panacea. The limits aren't well defined. Still, in the case of self driving cars, deep learning is state of the art. Nobody is hand-coding rules for how to drive the car. It's all about having as much good data as you can gather.
That's a luxury problem. It's like saying that you're better off without ECC because you don't have to deal with memory corruption. With a single sensor, you still have to deal with inaccuracies, the difference is that with only one you just don't know if you have it or not.
complexity is extra-linear, so you have to have to do the cost-benefit analysis to determine if it's worth it. and it may well be, but my point was that it's not a knee-jerk answer to always add more sensors. there is clear benefit, but is it worth the cost?
Not to mention that one should always strive to avoid a single point of failure. A "single reliable data source" is - whaddya know - a single point of failure. If that data source stops being "reliable", then you're screwed. When you're screwed in a giant hunk of metal rolling down a highway at 65+ miles per hour, then the only system you have left is to pray to your god(s) for a good afterlife.
To me as a layman, Car AI really seems like one of the most "intelligence" type of AI. Yes Alpha Go and Deep Blue are really good at something that was once thought "easy" for humans and hard for computers, but AI cars seems like trying to construct part of an animal's nervous system and sticking it into a vehicular body.
Radiology, and healthcare overall, is set to be improved rapidly if we can get some more labeled data. The hardware and software is already ready to start learning how to diagnose cancer and other health issues based on human-scan data.
Not sure whether an "intelligence" rating is possible, given that AI systems are still designed to be domain-specific, i.e., self driving cars aren't detecting cancer.
This just underlines what AI actually means and what is currently trying to pass for AI.
Isn't the point of SDC that they are better than humans? Shouldn't we support a diverse array of sensors to cover all the eventualities that we can? If some form of radar + camera solves this problem, what's the point of arguing humans would get confused by it too?
To be clear, the article alludes to the existence of three types of cars in the wild,
(1) Cars with many types of sensors (2) Cars with some types of sensors (3) Cars with no sensors
All other things being equal, why are people voting for (2) rather than (1)?
First you try countermeasures, if that fails you remove yourself from traffic. At least that's what humans do in that situation.
>or - when photons cover the subject that the camera/sensor is supposed to be reading -- last this happened the rider died
A frequent problem for human drivers too.
We just need good heuristics how to come to a safe stop, and be generally better than the average driver.
Not sure how well these things are currently managed by self-driving cars??
This sub-title is without context. Would be nice if the title here matched the article title, "This Image Is Why Self-Driving Cars Come Loaded with Many Types of Sensors"