As someone who has had family cars run into 3 times in 1 year at stop lights, by folks looking down at their cell phones.
I'm sure that Tesla's autopilot is already better than human drivers.
Looking out of my office window onto a busy intersection with bad visibility, that's easy to believe. It's insane how many accidents happen solely because drivers are unwilling to slow down for a second.
Hmm is he referring to the ability released to the public to date or their capability in general that isn't yet accessible by the general public? If it's the former that doesn't make sense so I'm going to assume the latter.
Considering in 2014 there were 32,675 motor vehicle deaths, 2,300,000 motor vehicle related injuries and 6,100,000 reported accidents in general (which means a likely higher number of accidents though not sure by how much) [1] it would not surprise me that even buggy software could do a better job than humans who tend to become distracted with almost anything. Though I'm not sure about extreme conditions like snow, heavy fog and rain, etc.
As much as I love driving (and I really do) I can't wait for autonomous vehicles. Considering it's a, if I remember correctly, a 1.7% chance in your lifetime that you'll be killed in an auto accident in the United States (and higher of a chance that you'll known someone killed) that's just way to high of a number for me (especially being a Dad).
It is worth keeping in mind that even without automatic vehicles, vehicle deaths have and will continue to fall massively.
From the worst in 1937 (29.357 per 100K pop) to today (10.345 per 100K pop), and falling by on average 2%/year since 1996 (was 15.8599 in 1996).
Without trying to sound melodramatic, cars built in the 1990s are "death traps" by 2015 standards (and the same is true of 19980s cars by 1990s standards). The reason we're seeing it fall slowly is as the late 1990s vehicles gradually disappear.
I suspect we'll see a massive drop again in the next ten years even without fully automatic cars as systems like Eyesight auto-break to slow collision speeds.
One thing which will get interesting when we do have completely automatic vehicles is: current drink driving laws (in particular in the US) aren't compatible with them. In many US states falling asleep in a parked vehicle with the keys is enough to get someone for "drink driving" (even without the driving part), so falling asleep in an automatic car could still get someone a DUI.
Thinking about your last sentence: Elon has been tweeting about Tesla's summon feature - tap your phone and your car will come to you, across the street or across the country. If a car can drive LA to NY legally, without a driver, there should be a way for it to carry a drunk passenger. Whether that means enabling "drunk mode" so you can't grab the steering wheel, or that you have to sit in the back seat, there's surely room for a compromise between the automakers and lawmakers. Technology like this will save lives.
There is plenty of tech out there that can "save lives" but the market has steadfastly refused. How about "No seatbelt, no car move". They tried, but given how many americans refuse to wear seatbelts it has never made it into a product. An enforced 'Drunk mode', presumably with some sort of breathalizer interlock, would be far more disenfranchising and I have to believe would also be rejected.
I like the concept but what if the sensor fails? Suddenly you're stuck who knows where...
I believe Chevrolet has a "teenage driver mode" which allows you to disable the radio, AC, and for it to bug you if you drive without a seat belt. I like that concept far more since it doesn't leave you stranded because the seat belt failed to be detected. It also allows you to set a maximum speed, a maximum radio volume, and for parents to get a report later if they wish.
I never drive without a seat belt. I just don't want to add more points of failure to a car than it absolutely needs.
Wait for the day that your google car won't move because it cannot connect to the server. Or because the lidar lens is dirty. Or a tire pressure sensor is incorrectly indicating a flat. Or (and I really hope this does happen) the engine light has been on for a week. Autodrives will no doubt have many reasons not to move. Seatbelt mandates will seem trivial in comparison.
Because regular cars never break down? I think with self-driving cars it would be that much easier for a remote health check to figure out something is wrong with your car and send you a loaner, analogous to the way tesla lends people cars when anything goes wrong with theirs.
An autodrive call will break down as much as a regular car AND then will brake down more because of the various systems over and above that a regular car does not rely upon.
When a regular car brakes down there is normally something you can do. The vehicle is still movable. An autodrive car with a blown lidar unit is probably bricked.
> An autodrive call will break down as much as a regular car AND then will brake down more because of the various systems over and above that a regular car does not rely upon.
I'm sure glad you have the data to back this up.
> When a regular car brakes down there is normally something you can do. The vehicle is still movable. An autodrive car with a blown lidar unit is probably bricked.
You can still put it on the back of a truck or tow it, the same way you would for a car which has actually broken down. You can do the same for the theoretically broken replacement car.
(1) Are you kidding me? This is basic logic. Adding additional failure points to a system, with everything else being the same, creates higher failure rates.
(2) If you need a truck to move it, I call that bricked. Normal cars, when they loose various systems, are still mobile. A normal car can still be rolled even without any electrical system working. Steering and brakes can be are completely controllable via muscles alone, even on electric cars. An autodrive car could be a total unknown. Some don't even have a manual (non-electic) steering wheel.
I was thinking more of something like "I'm drunk, so I'm going to voluntarily give my car control to take me home." Basically turning it into a driverless taxi.
"No seatbelt, no car move" is restricting in that it's always enabled. Something like what I've dubbed "drunk mode" or "taxi mode" gives you more options.
So a drunk mode that need the drunk to acknowledge that he/she is drunk and voluntarily initiate drunk mode? I don't think drunks are the best judges of whether they are drunk, or their driving ability. The law agrees. For drunk mode to be safe, to properly make a dent in the drunk driving problem, it has to actually prevent them from taking the wheel even when they want to, especially when they want to. Otherwise all the confident drunks, the alcoholics, wil keep taking to the road.
This is important because so many of the "human factor" mistakes cited by autodrive proponents actually begin with alcohol/drug use. If the human really cannot be trusted, the car has to take away his authority.
Really, what has changed about cars, besides automatic breaking? I agree that that feature saves lives. But I wonder about other stuff. I have a '99 Accord :) (which I barely drive)
> Really, what has changed about cars, besides automatic breaking?
Great question.
- Electronic stability control (ESC) (required from 2009).
- Anti-lock braking system (ABS) (in most cars from 2004 due to EU regs, although not required in the US until 2011).
- From 2003 the IIHS started rating vehicles for: rollover risk, side impact, and frontal offset impact. So car manufacturers started improving their designs to score better in these tests. The whole "top safety pick" thing is an IIHS rating.
- From 2009 the NHTSA required a greater roof-crush load (3x the vehicle's weight, from 1.5x).
- Safety Cell improvements: Crumple zones are less likely to trap in the driver or front passenger's legs/feet during impact.
- Anti-intrusion bars: Are more standard. The IIHS tests kind of pushed them into mainstream.
- Additional air bags (side impact).
- Whiplash reduction seats and headrests.
Overall regulations aren't why cars are safer in 2016, consumer demand is. People decided they didn't want to compromise on safety, so as a safer and safer vehicles became available consumers flocked to them. Manufacturers had a financial interest in not only be safe but the absolute safest in the road, and that competition pushed vehicle safety forward.
IIHS also started the Small Overlap testing a few years ago which was pretty nasty and lots of existing cars didn't qualify. It was a major factor in us purchasing new instead of used and the particular cars we went with.
Indeed. It is scary how badly cars were performing in their small overlap tests.
Just for clarity, they did "frontal offset impact" tests a long time ago, small overlap however was a recent addition, you can see the difference between the two here[0].
So even in the early 2000s vehicles did improve for frontal offset impacts, they just improved further in the last couple of years with the small overlap tests.
Switching to stronger and lighter materials has certainly helped. There's new tech like blindspot monitoring and backup cameras in almost every vehicle now, and even rather simple stuff that's been around for awhile (like daytime running lights and having a center brake light) makes a huge difference.
Thousands of tiny improvements to how cars crumple in crash tests. It's all incremental.
One change you can see on the road: the new looong Jeeps. That's so the gas tank can be ahead of and protected in a crash by the rear axle, a safety improvement.
> From the worst in 1937 (29.357 per 100K pop) to today (10.345 per 100K pop), and falling by on average 2%/year since 1996 (was 15.8599 in 1996).
Yet, the statistic is still twice to thrice as high as for other car-heavy nations like Germany or the UK, no matter whether you go by incidents per pop, incidents per vehicles or incidence per miles driven.
There's a lot of room for improvement outside technical gadgets.
I definitely think that systems like auto pilot will be better than human drivers, but that being said, I don't believe he has enough data to make the case for it being better TODAY.
Musk has a habit of going off half cocked, and I think this is one of those cases. Auto pilot is great, but you cannot say much difinitively about it because it hasn't got enough miles or road conditions under its belt.
That all being said, I myself have invested in this type of technology (Subaru's Eyesight) but auto pilot claims to be able to do much more than the more stabilised technology found on mass market vehicles.
Well, if that's the case, then why are humans crashing into others in prefect conditions? It's data that is fair but for certain conditions, as you said. I am interested in how they handle in extreme conditions but the reality is that it is FAR safer in mild conditions BECAUSE they aren't distracted by their cell phones or the sights or being tired/drunk, etc.
When they talk about "Tesla tweaks the software running its vehicles by analyzing data from the hundreds of millions of miles driven by current owners."...
Does that mean that they analyze the miles driven using autopilot or the miles driven by the human driver or both?
It would seem that using data from human driven miles would give them good insight into how human drivers would react in weird and non-standard situations.
Also, with as many cars as Tesla has out on the road, I wonder if the amount of data they have to feed into their self-driving systems is more than Google or other players in the space.
I was just imagining an autopilot lane with cars flying into the city to pick up their owners.
You could also have parking garages with a much higher capacity per sq. foot if all the cars were smart and could work together. (Or even sections designed for this) Instead of single or double car rows, you could have them 4 or five cars deep, and self-rearranging to let each other out.
Better perhaps a very specific tasks. I'd still like to see them respond to someone controlling traffic with hand gestures (a routine thing) or judge how best to get out of the way of an ambulance, which occasionally means breaking laws/rules.
Yeah, I think there's more to driving than just proceeding on the correct route and obeying laws,signals & signs. So I can see autopilot coming pretty soon but I think fully autonomous is much further out.
better in normal circumstances maybe. probably terrible in corner cases, which as he said, is where the real difficulties are and there are millions of corner cases.
It'd be interesting to know how many road accidents are those corner cases and how many are simple inattention. I've been in a couple accidents and each has been something I'm sure even a basic autopilot would've easily prevented.
1. You're right but you are thinking of human corner cases. AI also has corner cases which might be trivial for humans. I'm saying these corner cases for AI are too many currently.
2. And anyway, from my point of view 'better than humans' doesn't necessarily mean 'causes fewer accidents'. If that was the goal then an auto-pilot that kept the car still (or drove very very slowly) would win.
As someone who has had family cars run into 3 times in 1 year at stop lights, by folks looking down at their cell phones. I'm sure that Tesla's autopilot is already better than human drivers.