Hacker News new | ask | show | jobs
by the8472 3507 days ago
There seem to be some issues. E.g. it recognizes the joggers (0:55) as obstacles even though they're obviously (to a human) on the sidewalk. And when taking a right turn (1:01) it mistakes a hydrant and a parked car in a parking bay on the other side for an obstacle and stops for a moment before noticing that they're not actually in its lane.

I think a polite driver may have slowed down gradually to not startle the joggers, but not stopped if it were avoidable, especially since there was a car behind the tesla.

In the latter case i guess the field of view of the cameras is not wide enough to assess those things during a turn in advance?

6 comments

Interestingly, it behaved just like a young learner-driver would - play it safe, be cautious, and give extra space to the joggers because a new driver is a little uncertain about controlling a 6,000 lbs hunk of moving steel so close to delicate humans.

It's astounding to realize the AI will learn and change it's habits, just like a young human driver learning.

The car might have been playing on the safe side. If you look at the joggers carefully, only the left one is marked as an "in-path object", the other one has just an "object" box. They could probably tune down the collision prediction algorithm just a little bit, but that has to be done very carefully.
Why bother, err on the side of caution with pedestrians. The driver is already benefiting from auto pilot, a minor delay due to safety of other road users would be acceptable to most.
It's inconveniencing other drivers on the road and much more importantly this erratic stopping from a human perspective might make it more likely to cause an accident. Granted the accident would legally be the meat driven car's fault who bumps into the suddenly stopped Tesla. It would still be nice to play well with human drivers who don't have the luxury of relaxing in a self driving car.
> Granted the accident would legally be the meat driven car's fault who bumps into the suddenly stopped Tesla.

Exactly. If you bump into the car in front, because they stopped for a percieved hazard, it is your fault for tailgaiting, not their fault for percieving a hazard.

That's all very nice in theory but, in the actual real world, if you do a lot of unpredictable braking or other unexpected behaviors, you will basically get hit all the time even if those accidents aren't technically your fault.
Soon people will either adapt and end up following the law, or they will be unable to pay for insurance, because of the amount of claims that they're taking out against it.

I generally not a fan of social darwinism, but in this specific case, I think that something of the sort will have a positive effect. How many people have been injured because the car they were hit by was being forced to go faster, because of tailgaiting?

I agree, legal or not, I very much rely on other drivers "signals" when driving. Over the years you learn to predict what the driver in front of you would do in a normal situation and adjust your driving in advance. I'm still able to break on emergency but when you can clearly see way ahead of the car in front there is no obstacle and yet they still brake it's extremely risky and frustrating. Young drivers here usually mark the car some way to warn you to be super aware of their unexpected behavior, I'd expect self driving cars would to the same.
It means that you would have been surprised if the jogger was making a jump to her left (avoiding hole in her path for example), forcing any driver to hit the brake.
Of course you are correct. However, we don't need to make those incidents more frequent than then already are.
I wish there were some external visual cue like a light or something to indicate a car is using autopilot. That way you could give them a wider buffer.
Maybe a pulsing red light, moving back and forth horizontally.
Instead of relying on other drivers' signals, you could instead use them as extra warning, and instead of driving so close to them you're putting yourself and them and potentially others in danger if they brake when you didn't expect, you could back off to a safe distance.
Well, not sure where you read I drive close to anyone. What I'm trying to say is that erratic drivers are dangerous because you learn to predict what a normal driver would do in a typical situation (e.g. jogger too close to the lane => slow down to a safe speed and navigate towards the center).

If a self driving car doesn't reproduce the most common driving style the other drivers should be extra careful around it, some kind of mark on the outside could help.

Well no one think of the coffee?
The problem is that pedestrians on the sidewalk could potentially become an unavoidable obstacle at any time if they walk into traffic. For this reason there can never be a 100 percent safe system until vehicles and pedestrians are physically separated by barriers or separate structures which may happen eventually in some new urban designs.
the jogger is a problem because the legs are so flimsy :)

If you don't see the legs and don't have very accurate 3d vision, the object might just as well stand further away in the middle of the road: http://malea.lacerta.uberspace.de/up/e92e6492-ce56-468b-8e09...

I think the hydrant is the same problem: head on you don't have any parallax effect to do know the distance, so if you don't detect where the thing is rooted to the ground you might guess it's in the way.

and it completely fails do detect the right end of the lane (the moments before I was amazed how it still drove so fluid)

I think 0.25x is the right speed if you want to watch it in realtime

Does this new Telsa autopilot use stereo cameras? Would stereo camera help to have a better 3D vision? Or is a video stream enough to get the 3D data.

From the video it looks like lane detection and 2D object recognition.

It might use a structure-from-motion (SfM) running on CUDA to get a 3D point cloud out of the 8 cams. Does someone know more about how their new auto pilot works?

They do not. They have 3 forward cameras (1), but they all point directly forward and are for separate use cases. I've always thought it would make more sense to have two cameras at either side of the windshield rather than one in the middle. The downsides are you need a new method to clear water and debris since the wipers won't hit that area, and there might be limited FOV.

Stereo might require some level of calibration to account for camera alignment, maybe they want to avoid that? I would think that implementing stereo would be fairly cheap both computationally and in hardware- if they can get by this well without it, they could pare the stereo computations down to a few frames per second, on specific objects in the scene. Maybe a future update will add binocular cameras in the dash, but they are pretty confident about their current hardware.

1: https://www.tesla.com/en_EU/autopilot

I'd imagine they use stereo cameras. AFAIK though, their main sensor is radar.
I would guess that the car couldn't predict the direction the joggers were moving in. They could be jogging towards the road. We as humans would know that they were running along the road because we see their backs but not necessarily from extrapolating their movement
It passes a single jogger without an issue beforehand, but they were two when the car stopped and left one was relatively close to the lane when the car was approaching. Maybe it wouldn't stop if it was a single jogger.
If you pause it at 0:54 it clearly made a mistake. The foot of the left person is outside the green box, the car thought the person was further forward and to the left of where they really were. Good progress but not ready for prime time.
you're looking at it from the wrong perspective. the car just wanted to slow down to give the male driver time to properly look at the female joggers.
I don't want my car to be like the creepy telepathic dog from "A boy and his dog" that is always looking for partners for his owner to mate with.
i don't know much about self-driving, but i would not over-interpret the green boxes. Also i don't know if your assessment about the range is true. Seems like a bug.