Hacker News Reader

Comment by dbecker | Hacker News Reader

I'm trying to figure out if I can keep my current car until autonomous vehicles become broadly available. Anyone here have an informed opinion on when that will happen?

View on HN

Unless you live in an area with perfect weather, don't expect autonomous vehicles very soon. They have difficulty in heavy rain, snow, and ice.

Because we humans do not in the exact same circumstances.

In the exact same circumstances, humans can adapt to and handle the difficulty, whereas autonomous car technology currently can't and doesn't.

That's a significant problem. So does Google only send their car out in good weather?

The cars are developed in California. The weather is great here almost all the time. The cars are still under development and inclement weather is an obvious thing the team must be thinking about. I fully expect, in time, that autonomous vehicles to drive more safely in bad weather than people. (They'll be more cautious, for one.)

Absolutely. If you've ever driven though heavy snow, you know that you can barely see any of the road (no white or yellow lines) and lights are the only part of other cars that you can see. Signs and other traffic signals get completely covered with snow. People can barely drive in those conditions, I highly doubt that a self driving car could with current technology. Although I have high hopes for the future.

Why would machines be worse than humans in bad weather if they are better than humans in good weather? Its not like our eyes are better than the machines cameras.

If its because human drivers can better "build a mental model" of whats happening with limited visibility/data - that seems like the same problem that autonomous cars are designed to solve.

Whats stopping autonomous cars from making progress in the bad weather arena? Is it just we haven't done enough miles in that weather and we need to train the ML some more?

You're putting a lot of faith in the ability to solve a very difficult problem. Computer vision has made huge strides in recent years, but it's still far behind the capabilities of the human visual cortex. It's not just a matter of building a simple model and training it until it achieves arbitrarily high accuracy. ML encompasses a powerful set of tools, but it's not a magic black box that eats data and spits out AI.

We don't yet know how the human brain processes images at a detailed level, and it's too complex of a system to optimize by brute-force. The best ML image processing systems still rely on a lot of pre-encoded assumptions, and require huge amounts of training data/computational resources, and don't perform as well as humans even on simple tasks.

For now, autonomous vehicles "cheat" a lot by using sensor technologies like lidar and radar -- techniques that produce spatial data that can be interpreted more easily than raw imagery. That adds a lot to the hardware costs, but it's currently the only way we can make them perform acceptably well in real-world scenarios. And conditions that interfere with those technologies more severely, like fog or snow, are still crippling.

Actually there's no reason to suppose that computer vision hasn't already exceeded the human capacity. After all, it has lidar and radar. Cheating or not, it has a very good result.

Agreed -- don't knock it if it works! However, those kinds of sensors don't really fall into what's normally considered "computer vision".

Mind you, I'm extremely optimistic about the potential successes of autonomous cars in the near- to mid-future (the next 5 to 25 years, say). But I don't want people to get the wrong impression about where the state of the art is today. And in particular, I want to push back against the idea that machine learning is a solved problem, and that autonomous cars can teach themselves to drive as well as a human if we only give them enough miles of practice.

Vision algorithms are pretty shaky to begin with. We've made leaps and bounds of improvements but they're still pretty far behind what people are able to do. In inclement conditions I think humans still out perform machines pretty handily. Eventually it'll be an even field or there will be sensors added to cars that can just see through the snow/rain but that's not where we are.

> Why would machines be worse than humans in bad weather if they are better than humans in good weather?

[Because they are.](http://jalopnik.com/this-is-how-bad-self-driving-cars-suck-i...)

Basically, the sensors they tend to equipped with aren't even close to being sufficient in poor whether. Humans are EXTREMELY good at pattern recognition and can manage to make it work, but computers just aren't at that level yet. It'll take a combination of improved sensors and computer vision (ie, pattern recognition) to make it happen.

Because Google has mapped out the roads to a very specific level; this allows them to use sensors in the cars that don't have to have the fidelity of human eyes etc. As Google's maps become out of date, they need to be recreated or checked. Their system won't scale...

While heavy snow / rain is a hard scenario to get self-driving cars into, I would argue that they have the possibility to be better than humans. We are dependent on our poor quality vision. Cars could use laser, low light vision, thermal imaging, etc. to look right through heavy snow or fog!

>you can barely see any of the road (no white or yellow lines) and lights are the only part of other cars that you can see.

Arguably, no car should be driving in those conditions unless it is a life or death situation. Because you have a very good chance of dying doing so.

In some parts of the country, it's pretty normal. The roads are caked with snow for weeks at a time and you can get into white out conditions if the wind picks up. As long as you slow down accordingly, it's usually not an issue.

Then put your auto in manual mode and drive it by hand. No reason not to roll them out in the next five years. If it detects rain or snow it simply refuses to self-drive.

Perfect! I live in San Diego, CA, so I can expect my autonomous car soon.

If you're fine riding with other people, then a self driving car is just a Lyft Line with one extra seat available and without driver wages to pay. I think they're going to drive down the costs of a ride to a level where buying a car doesn't make sense for many people even before self driving cars happen.

~5 years. Freightliner's auto-drive truck is already in production, and light vehicle manufacturers, Google, Apple, Bosch are all pushing the limits pretty fast (and Tesla of course!).

5 years tops. I'd bet on it.

Me too. However I'd rather wait for a few years after they're out, like the Apple Watch, before I use one.

I think we won't need to own cars and will simply press a button like Uber and the car drives us wherever.

I'll (sadly) take the over on that.

I'll bet you $1k that fully autonomous cars will not be: (1) legal to ride on at least 90% of public roads AND (2) available for sale to the general public AND (3) for under $250,000 in 2015 dollars AND (4) in the state of California AND (5) on or before June 5 2020.

Deal? (If I were richer, I'd bet more, in order to up my chances of being happy either way. I want to lose, but don't think I will.)

I think we're going to quibble on 2 and 3. It doesn't matter what the capital cost of the vehicle is if you no longer need to own the vehicle and can order it on demand from Google, Uber, etc.

If you want to hammer out details, my email is in my profile. We'd use LongBets [1] to keep track, and under the condition that the loser has to donate the amount chosen to a charity of the winner's choice.

[1] http://longbets.org/

I don't see how autos could realistically cost $250k each anyway. They are not using significantly more raw resources. Electric cars are an accurate comparison, because they also have top of the line tech in a specific domain - batteries. And volume batteries are a much bigger manufacturing problem than radar systems, which have been assembled (but not at auto-fleet scale) for decades, albeit not with current tech, but much more closer to current assembly than what Teslas LI batteries are.

I imagine something like this Google car could go to market for around $80k today, and thats factoring in huge amounts of per-unit overhead for building factories to assemble them en masse in the first place.

An auto upgrade kit to a traditional car probably won't cost more than 20k if they permit the retrofitting of old vehicles when the ball starts rolling.

IMHO, fully autonomous cars for the general public are at least 10+ years away.

What I'm excited about is how we will soon see more and more autonomous 'features' i.e. lane changing, smarter adaptive cruise control and parking lot mode.

I'm hoping by 2025, but that may be a bit optimistic.

I suppose if they become mainstream, you will be able to convert your car to autonomous.

I don't think that a conversion will be easy. In addition to all of the necessary sensors and automated controls, you'd need to find someone willing to either do the conversion for you or sell you the parts needed to do it yourself, including a license for the software. Given the liability issues and low payoff involved, I doubt that any of the players involved will be licensing their software for diy/conversion use.

Isn't the long term plan of google to licence the technology to car makers? Also the sensors and manipulators seem like standard components so they 'll probably not cost a lot when they are produced at scale.

I believe that is their plan. However, I don't see them going after the aftermarket—too many variables, and given the liability issues involved (for instance, what would happen if a sensor was installed incorrectly?), I think they'll avoid the aftermarket channel entirely.

At the opening yes. Midgame, is requiring autonomous cars only in certain situations - perhaps highway/freeway lanes. Endgame is, of course, autonomous only - aftermarket is inevitable there. But the tech will be old and reliable by then.

I don't see us ever getting to a point where human drivers are prohibited—at least not in the time span where we'd need to worry about converting existing vehicles. In the USA at least, driving is so closely associated with the idea of "freedom" that such a move would be politically untenable, no matter the potential for lives saved. Also, I think that once the majority of vehicles are autonomous, the remaining human drivers become less of a danger—I'd argue that most accidents involve two parties making mistakes (even the "not-at-fault" driver wasn't driving as defensively as they could have been).

It will be interesting to see how insurance rates change for autonomous vehicles—a vehicle with no potential for manual control should in theory be much cheaper to insure.

Reply on news.ycombinator.com