3 ms·
> Automatic braking doesn't get much attention, but it seems like something that could be incredibly valuable if it got to the "clearly better than human" level
by ajross 5y ago
> Automatic braking doesn't get much attention, but it seems like something that could be incredibly valuable if it got to the "clearly better than human" level.
It's important to point out here that a handful of back of the envelope analyses are out there that show that Tesla's autopilot is already "clearly better than human" level at avoiding emergency vehicles.
These are very common accidents. Emergency vehicles in travel lanes get hit all the time, crews are carefully trained to avoid that situation if at all practical, and taught how to reduce personal risk when that's not possible.
In fact (and I don't have a link handy) if you extrapolate out from a few different data sources it looks like Tesla's on AP get in these crashes about half as much as vehicles in the general fleet.
Now, that doesn't mean that these 13 events don't constitute a cluster worth investigating. And it doesn't mean that there can't be a bug worth fixing. It doesn't even mean that the media shouldn't write about it.
It does mean that we should be very careful with pronouncements about safety from a sample set of this size, though.
- lacker 5y agoI think the problem here is calling something "better than human" because it manages to avoid some accidents. If a car avoids one accident for every 100,000 times it intervenes, that isn't really a good thing. It just means that drivers start to ignore flashing lights and warnings because they are always meaningless. Maybe it even causes some accidents by distracting the driver, and they don't get categorized as "caused by the automatic braking system" because nothing really tracks that. Every time I start my car, about 15 seconds into the drive a popup appears on the screen, saying something like "Warning: automated driving assistance system can cause distraction. Select OK to dismiss". Thanks, you have reduced your own liability in exchange for further distracting the driver.
- syrrim 5y ago>In fact (and I don't have a link handy) if you extrapolate out from a few different data sources it looks like Tesla's on AP get in these crashes about half as much as vehicles in the general fleet. Why use that as a comparison? The obvious confounder would be that people would disable autopilot in poor conditions, making such crashes a priori more likely. The relevant comparison would be Teslas (with and without AP) against a similar class of cars, maybe BMWs.
- ajross 5y agoThat's nitpicking. There's an initial contention: "Teslas hit emergency vehicles frequently", and a simple refutation: "No, they don't, here are some numbers[1]". Your response only makes it "not impossible" that the initial contention is correct, but you have to show numbers to prove that. All we have is a cluster of 13 accidents over 3-4 years, and there are something like thousands of such accidents in the US every year. That's not proof that this is noise and not signal, but the obvious hypothesis is that it's probably noise[2] [1] Which I don't have, though I suppose I could track it down. [2] For the record: I believe it's probably a real bug, but in a product that on the whole is clearly better, so it's well hidden in the noise.
- syrrim 5y agoThe initial contention was that tesla's autopilot was superior to humans, at a specific task. If this were true, it would be very interesting, and represent real progress in the self driving space. Thus I'm interested in whether it's true, and proposed a method by which that might actually be determined.
- emn13 5y agoWhat's the source for this claim? Are you accounting for the fact different classes of vehicle have differing accident statistics (e.g. comparing teslas to similarly old and expensive alternatives), and that autopilot's miles are not a random sample of all driven miles?
- ajross 5y agoNo, because it's enough that it exists as an existence proof refuting the hypothesis. The burden is on people contending that e.g. AP can't stop for emergency vehicles (or the related point in the linked article that emergency braking doesn't work) to show that it is unsafe. That analysis doesn't hold for obvious interpretations of the data we have. That doesn't mean it's wrong, it means you need to come to the table with more work. It's not enough to argue with me on HN.
- emn13 5y agoIf you're merely questioning whether Tesla's really do crash into emergency vehicles more often that would be expected, I get it. But "Tesla's on AP get in these crashes about half as much as vehicles in the general fleet" is unsupported; even in Tesla's own figures from https://www.tesla.com/VehicleSafetyReport https://www.tesla.com/VehicleSafetyReport the difference between AP and non-AP Tesla's isn't large enough (for me anyhow) to actually draw any conclusions from, since AP usage is obviously correlated with road conditions. To be clear: I'm not saying it's at all obvious that AP does worse than a human driver, but it's neither is that possibility ruled out by anything they've posted there - Tesla simply choose to present statistics without the context required to interpret them, likely because it suits their marketing aims. News at 11. The human instinct to retain control even when that's not to their advantage gives me the hunch that it's more plausible that the apparent numerical benefits of AP are simply due to selection bias; people turn it on in situations that are less messy, and therefore it's involved in fewer accidents. Perhaps these investigations will reveal more. Put it this way: if autopilot actually were better than a human driver, and humans choose to use it in less dangerous situations, and those drivers at least sometimes pay enough attention to catch the occasional AP oversight (better isn't perfect!) - well, then I'd expect the AP crash numbers to be dramatically lower than the human crash numbers. They're not. I do think the Tesla AP numbers look less bad than they did a few years ago anyhow, to be fair. I just don't think it paints a clear picture yet.