Tesla’s Autopilot is suspected in the fatal rear endings of two motorcyclists. With deep learning doing automated feature extraction, it's impossible to know...
Anecdotally, I’ve been on the full-vision stack for several months now and it’s significantly better at detecting pedestrians and motorcycles. I’ve also had no phantom braking events since I got into the beta, which is just as important to me. Just last night I had it drive me 10 miles home on Pennsylvania back roads and I didn’t have to touch the wheel until it got to my house.
That said, I think having a detailed 3D map of the environment directly out of the sensors is far more useful than trying to generate one from video, but I also don’t see a reason not to do both. I’m pretty sure they use Lidar to train the FSD model, too.
I suppose he is betting on the public accepting these deaths as non-important glitches, and no regulation coming for Tesla. So in a very cynical sense, I suppose this profit gamble is playing out for Musk.
However, how come other car manufacturers using lidar aren’t making a point out of this? It can’t be out of good will, surely.
Not to be dismissive of these deaths, but there should not be any expectation of self-drive cars being perfect, ever. The reality is that if they are safer than humans overall, then they have reached a point where we can (and should?) adopt their use. It’s not a huge surprise that there is some form of bias in the current deaths simply because biases arise in any complex, real-world systems.
We should, and must, accept some glitches.
But, uh, yeah Tesla may well try to delay addressing the biases if we don’t call them out, so this info is good.
I agree, these systems are fundamentally a very good thing and the more cars they’re in, the safer we are on the road.
I just don’t think they’ll be substituting human beings at the wheel of cars in the near future, and I personally think it’s not so good of an idea to frame it like that.
I mean, we’ve only automated some, not all metros all over the world, and no passenger trains at all. If we haven’t figured that out yet, I can’t see how self-driving cars are supposedly just around the corner.
Because there’s more to this than “bad car kill people.” Two deaths are tragic but miniscule in the context of daily road fatalities which headlines like this never include. Nor do they include stats on cars with driving aids such as these and how they’re involved in accidents 1/10th as often as typical cars saving many more lives in the process.
Also, these cars had someone behind the wheel who should have been watching the road but were too busy playing on their phones or whatever when they hit someone, same as would happen had they been in a 1995 Honda Accord.
Sure we’re talking about very few deaths, but it still is a design flaw, as it has trouble recognizing a specific kind of motorcycle. I would say that makes it more of a bug that hasn’t been patched out rather than a mere statistical error, but I am not well versed in actual software development so someone else might come up with a better analysis.
As for human input, I agree, this is a very different case from, let’s say, an experimental self-driving car. However I still believe Tesla’s decisions play a part, for example the naming scheme they chose, autopilot rather than assisted driving or some other admittedly less enticing name.
Of course one might say that people with a car license should be able to see through basic marketing, but it might nonetheless influence people’s behaviour, even if just subconsciously.
Because Musk’s rejection of Lidar was foolish
That’s one of the points brought up in the video.
He didn’t just reject it, he condemned it for everybody in the industry. What an asshole.
why? what he said about lidars?
https://youtu.be/yRdzIs4FJJg?t=349
That lidar systems are doomed to failure.
Super long term when machine vision works very well he’s right. The question is how long will that take. 5 years? 20 years? 40 years?
Anecdotally, I’ve been on the full-vision stack for several months now and it’s significantly better at detecting pedestrians and motorcycles. I’ve also had no phantom braking events since I got into the beta, which is just as important to me. Just last night I had it drive me 10 miles home on Pennsylvania back roads and I didn’t have to touch the wheel until it got to my house.
That said, I think having a detailed 3D map of the environment directly out of the sensors is far more useful than trying to generate one from video, but I also don’t see a reason not to do both. I’m pretty sure they use Lidar to train the FSD model, too.
Maybe he should have invested in more AI instead of buying Twitter to ban people.
it was a good profit decision
I suppose he is betting on the public accepting these deaths as non-important glitches, and no regulation coming for Tesla. So in a very cynical sense, I suppose this profit gamble is playing out for Musk.
However, how come other car manufacturers using lidar aren’t making a point out of this? It can’t be out of good will, surely.
Not to be dismissive of these deaths, but there should not be any expectation of self-drive cars being perfect, ever. The reality is that if they are safer than humans overall, then they have reached a point where we can (and should?) adopt their use. It’s not a huge surprise that there is some form of bias in the current deaths simply because biases arise in any complex, real-world systems.
We should, and must, accept some glitches.
But, uh, yeah Tesla may well try to delay addressing the biases if we don’t call them out, so this info is good.
I agree, these systems are fundamentally a very good thing and the more cars they’re in, the safer we are on the road.
I just don’t think they’ll be substituting human beings at the wheel of cars in the near future, and I personally think it’s not so good of an idea to frame it like that.
I mean, we’ve only automated some, not all metros all over the world, and no passenger trains at all. If we haven’t figured that out yet, I can’t see how self-driving cars are supposedly just around the corner.
Because there’s more to this than “bad car kill people.” Two deaths are tragic but miniscule in the context of daily road fatalities which headlines like this never include. Nor do they include stats on cars with driving aids such as these and how they’re involved in accidents 1/10th as often as typical cars saving many more lives in the process.
Also, these cars had someone behind the wheel who should have been watching the road but were too busy playing on their phones or whatever when they hit someone, same as would happen had they been in a 1995 Honda Accord.
Sure we’re talking about very few deaths, but it still is a design flaw, as it has trouble recognizing a specific kind of motorcycle. I would say that makes it more of a bug that hasn’t been patched out rather than a mere statistical error, but I am not well versed in actual software development so someone else might come up with a better analysis.
As for human input, I agree, this is a very different case from, let’s say, an experimental self-driving car. However I still believe Tesla’s decisions play a part, for example the naming scheme they chose, autopilot rather than assisted driving or some other admittedly less enticing name.
Of course one might say that people with a car license should be able to see through basic marketing, but it might nonetheless influence people’s behaviour, even if just subconsciously.
It’s also a lot of data to deal with, depending on resolution.