Tesla Autopilot Hits a Deer [video]
youtube.com
youtube.com
* "My Tesla was on autopilot and it plowed into a deer in my lane, there were no cars behind me, it did not attempt to slow down or stop."
* "As one Tesla autopilot mile becomes 50, 500, 5000, the behavior of your brain changes - it offloads the cognitive burden of driving to the autopilot, even when you have your hands on the wheel looking straight ahead"
* "Your brain cannot turn off habituation, it will put too much trust in autopilot, you can't do anything about that and that can be disastrous"
* Mentioned article: Tesla Drivers Using Autopilot Watch the Road Less, MIT Study Shows - https://www.caranddriver.com/news/a37727214/tesla-drivers-au...
* The footage when it hits a deer is at the 3:50 timestamp - https://youtu.be/jrWH_0YA5XM?t=230
> The person in the driver’s seat is only there for legal reasons. He is not doing anything. The car is driving itself.
[1] https://www.tesla.com/videos/autopilot-self-driving-hardware...
Most planes today can land themselves, yet pilots still often land manually to keep their attention and skills sharp. Tesla drivers which make use of Autopilot should consider doing the same -- since ultimately it's their ass on the line for now.
I highly doubt Tesla will start putting drivers through several hundreds of hours of training every year (let alone pay them 6/7 figure salaries) to have properly safe and trained drivers in their cars.
>* "Your brain cannot turn off habituation'
The entire argument falls apart, and quickly. If the statement 'its the drivers responsibility' relies on us completely ignoring human behaviours and patterns.
People are falliable, and encouraging people to not concentrate and put trust into auto-pilot, and knowing with every kilometre travelled. People concentrate less and less. Who is at fault when the car misbehaves and breaks away from expected behaviours?
So when Tesla says[1]:
> The person in the driver’s seat is only there for legal reasons. He is not doing anything. The car is driving itself.
The "legal reason" for having a person in the driver seat is so that they can act as a liability shield for Tesla?
[1] https://www.tesla.com/videos/autopilot-self-driving-hardware...
That is a specific marketing video from five years ago. The current system has basically nothing in common with what is shown there, hardware or software wise.
> The person in the driver’s seat is only there for legal reasons. He is not doing anything. The car is driving itself.
Does it say that in the Autopilot terms of use? Maybe the owner's manual?
No, I made two distinctively different comments, but used the same quote from Tesla in them, which is the same quote you've just quoted yourself.
> That is a specific marketing video from five years ago. The current system has basically nothing in common with what is shown there, hardware or software wise.
So it got worse over the last 5 years? Because Tesla said the car was driving itself, but now you're saying that's not true, and there's nothing in common with them.
Selling FSD however is pure scam.
Cruise control only maintains velocity, lane assist still requires interventions when turns need to be made or lines are unclear. That sounds like intervention to me
"without the intervention of the pilot" is a pretty pivotal part of everyone else's understanding of the word.
1. The feature marketed as "autopilot" on airliners is accurately named
2. The human pilots of airliners need to keep paying attention even when autopilot is active
Isn't that exactly analogous to how it works in a Tesla? Why do you then say "it definitely isn't that"?
To me it seems likely that putting in HUD-type assistance would be safer than attempting self driving, which is more complicated than landing a plane.
For landing existing CAT III landing will complete the entire landing and rollout. Failures (e.g. loss of a flight computer or sensor problems) that would lock out CAT III autoland would likewise rule out attempts in a HUD.
The reason you don't see much real CAT IIIc bad weather landing isn't that it can't work or HUDs would be better, the problem is that sure, now you're safely on the ground but with zero visibility you're stuck on a runway, you can't very well taxi a jetliner when you can't see the concrete you're on let alone the signs and lights. Even a follow-me car won't help, you can't see it.
Garmin's emergency autoland for GA pilots has the same problem, but since it's an emergency feature "Oh dear we took an entire airport out of service for an hour" is an acceptable consequence. The alternative may well have been fatal in that case.
Anyway, people have added e.g. enhanced visibility and highway edge marker HUDs to high end vehicles before. But, that's going to be something you need to train with in order to use it effectively. An airline would send crew on the recommended two week course before they're allowed to fly the modified plane, but we know even private pilots routinely avoid recommended manufacturer training, and cars would definitely be worse. My guess: Tesla would sell a thousand cars with a HUD, fifty drivers would take the free training, ten of them would come away having learnt anything. All one thousand owners would be confident they can now drive at 60mph in dense fog even though the training said "You definitely must not drive at 60mph in dense fog".
That leaves it to consumers and regulators. As consumers we can stop getting caught up in Musk's persona, and stop buying a dangerous product. And regulators should start doing their job for the first time in many years.
You're describing a flawed overall system. It has to be nearly perfect or it would be better not to exist at all. Middle ground is incompatible with our minds. You are asking for the impossible.
Or would that just train us to ignore the ‘take over tone’?
Like in this particular case - it was on a bend, there was an oncoming car approaching - should the autopilot have hit the brakes at full and risk swerving into the oncoming car? Should it have swerved around the deer and hit the oncoming car? I would have probably slowed down a bit but still hit it as human reaction speed is not that fast in reality. The deer becomes clearly visible ~3:53 and contact is made at ~3:56. Studies[1] show that the average reaction speed is 1.5 seconds. This leaves 1.5 seconds to come to a full stop. Would you be able to come to a dead stop within that timeframe at 70mph?
Every time a Tesla thread comes up I see so many "couch racers" in comments that pretend they're formula 1 drivers with split second reactions.
In reality this is a life-saving technology - imagine how much death, loss and suffering will self-driving cars prevent by taking drunk people off the road alone.
[1] https://roads-waterways.transport.nsw.gov.au/saferroadsnsw/s...
But she says the Tesla didn't brake at all, which suggests it just didn't recognize the deer or doesn't have a rule saying that an animal walking toward the road might be dangerous.
By your logic, one needs to move 30mph on the interstate to make sure they can react to whatever jumps out at them. Or am I misunderstanding something here?
I've encountered many situations like this without hitting the animal. I would slow down and move aside to give the deer room, starting as soon as I saw the deer at 3:53. If the deer had moved faster and taken the whole lane I'd have passed it in the other lane. The oncoming car wasn't that close.
But you must also be able to stop in this distance, because the obstacle might not be a deer you can steer around, it might be a fallen tree that blocks the whole road.
You're certainly right that many human drivers would not do better than the Tesla here, but we should expect the Tesla to be better than a bad human driver.
The speed limit is supposed to be the maximum appropriate speed in ideal conditions. In practice almost no one treats it that way, but that's the philosophy by which it was designed. They take into account the clear area on the side of the road, sight lines, and the curve of the road itself.
Deer standing safely by the road side will happily panic and jump in front of your car at the last possible moment. In theory there is no difference between theory and practice, but in practice there is.
If a child jumped in front of my car, I'd swerve even if that meant risking a tree. If I ran over a child I don't know that I'd be able to live with myself anyway, so the choice is clear: I risk myself to improve the odds of the child making it away unscathed. But a deer? No way in hell. I'll slam on the brakes, since my car has ABS, but I would not take my chances with swerving.
The car could have reacted better and likely prevented the hit, but if a human had been the only driver nobody would have even suggested that the driver was at fault.
When there are other cars around or when the animal would've jumped in front of the car before you could react, I'd say the driver can't be blamed for hitting the animal because there was no reasonably alternative. In this case, breaking and turning on the hazards would've been a perfectly fine solution. This deer didn't have to die and a human driver could've (should've) avoided a collision.
The deer is just an example of the system failing, though. It shows that Tesla's wannabe-self driving is incapable of dealing with animals near the road. A small deer like this could be hit safely, but when a car doesn't try to break for something bigger and heavier, people could easily die. You run over birds and bunnies, but don't even try to drive through a horse. That thing is going straight through your windshield when you hit it, so you'd better risk getting rear-ended when a runaway horse walks onto the road. You're going to need one of those death mobiles like a Cybertruck that doesn't do those fancy "crumple zones" or "safety glass" to make it through anything larger than an adult boar without risking a crash, so any full self-driving system should be able to anticipate large animals near the road (or shut down in areas where large animals are known to cross roads).
Well not a tesla. You have to really work to flip them, but the point stands.
This despite being taught about danger of animals, and living in a place with a large wild population.
Your brain does not process images outside of the immediate distance in front of you nearly as fast as you think it does. I had a good four seconds that I could have "seen" the dear, but it was only when it was directly in my path that I realized what was going on.
When you are watching youtube, you are staring right at it. When you are driving a car, chances are you looking straight ahead.
No. You should brake. The intended and authorised safety regime for automobiles assumes you will maintain a situation where you can reasonably brake before foreseeable impacts.
This is quite different from say, an airliner or a railway train which couldn't possibly brake, and so those need completely different safety environments. "Absolute block" is an example for railways, there is intended to be no possible way for two trains to be in the same section of track, so, you always have at least an entire section (it will often be more on fast trains) to brake even if you can't see far at all.
In a car you don't have that environment, you should always be prepared to stop in the distance you can see. If you're thinking, "But nobody drives like that" then you're part way to understanding why so remarkably many people die on roads.
With large farm animals such as horses and cows, or with larger wildlife like moose and elk, your life is at stake - there are plenty of drivers killed just hitting moose (look at the size of the mother in the embedded video and you can easily imagine her sliding up the hood and through the windshield: https://www.adfg.alaska.gov/index.cfm?adfg=livewith.drivingm.... This video also shows something that's true of elk and deer as well - they have very little traction on pavement.
2) The difference between the nose of a car while braking and while coasting is an inch or two. Won’t make a lick of difference.
Source for 1) Having a deer jump over my car, leaving only fur in the windshield wiper blades. 2) Mythbusters, measuring for a “Wanted” car jump myth.
Ironically I did actually end up having a encounter with a deer. I did break, and swerve, but at the time I had a car without anti-lock breaks. I lost control of the vehicle and flipped it (which is also why they tell you not to swerve, unless you have good awareness).
Of course, they also said to break as much as possible but hit it straight on. You would total the car, but cars are designed for that, versus your body not being able to take a deer going through the windshield if you clip it on the side. The engine block is padding, use it.
Looking at the video, it looks exactly like the videos of horrible car crashes I saw in school (less blood). As usual with Tesla, the headlines are sensational, the hacker news comments breathless, but it's a fairly straight forward physics and recognition problem. 300+ people die a year, and billions of dollars in cars are destroyed each year. A neural network mimics the way human brains work. If a animal goes in a unexpected direction (or if it wasn't recognized at all), its going to be a problem.
The headline is "Tesla Autopilot Hits a Deer" which is as bland and unsensational as you can get while being true.
> A neural network mimics the way human brains work.
This is like saying a drawing of a circle mimics the way the sun works. NNs are vaguely inspired by a high-level neuron topology, but the similarities pretty much end there. Gradient descent and backpropogation don't take place in the brain at all, and we don't know enough about the brain to mimic the way it works.
Sounds like a very dangerous robo-taxi if that was going to be released as "Level 5" for 2020.
https://www.youtube.com/watch?v=q0_eKr8jnNo
@2:20 exact same situation as in Jessa Jones clip. "all of a sudden a deer ran out onto the road but autopilot didn't react at all, gracie's applied the brakes only moments before impact but it was too late " Living creature hit, $4500 bill an 2 weeks without the car.
@6:30 plows into road debris roughly size and shape of a human lying on the road.
@9:00 finally correctly recognizes human on a crosswalk.
@11:00 Coyote, zero reaction and human takes over to avoid accident.
[1] https://www.nytimes.com/interactive/2018/03/20/us/self-drivi...
We try to make it some magic AI system, but instead it is dumb system that keeps speed and route potentially even between lines and it might even make turns. But that is pretty much it.
Now, FSD marketed by Tesla, should be something the whole company is punished with massive daily fines until they make it actually work or stop using the term and allow anyone to return vehicle for full price.
No production car (AFAIK) uses LIDAR. Some have RADAR, but it doesn't usually help a lot here. This Tesla actually has radar in it. Tesla is actually moving to pure vision because they think they can respond faster to a all vision system and have fewer conflicts when radar randomly bounces back (this apparently was one of the reasons for phantom braking events in autopilot), but also apparently because radar interference with other cars is starting to become a thing.
Various brands are putting LiDAR on their cars:
- Lucid: https://www.autocar.co.uk/car-news/new-cars/lucid-air-1065bh...
- Toyota and Lexus: https://global.toyota/en/newsroom/corporate/35063150.html
- XPeng: https://insideevs.com/news/537039/xpeng-p5-lidar-video/
- Volvo: https://www.cnet.com/roadshow/news/electric-volvo-xc90-lidar...
But boy, we do need that technology to come down in price, stat.
Knowing what I know about computer vision (which is quite a bit), this is actually one of the simpler problems they have to solve, which is why I think it'll be solved and what you're seeing here is just a data bug. Not enough animals in the training set.
Then what's the point of self driving? Thrills? No one would want it.
My Honda does something similar with its lane-centering feature (it will disengage after maybe 15 seconds of no driver input), but that feature is so limited it's not annoying.
Maybe limit autopilot so that for every mile, or hour spent on autopilot you have to drive an hour manually? Make it random?
Also stop calling it autopilot because it implies self driving capabilities that are not yet fully baked.