4 ms·

The villain has a point: "The reason why nobody has achieved this level of functionality is because today’s software is not good enough to predict the future. I

The villain has a point: "The reason why nobody has achieved this level of functionality is because today’s software is not good enough to predict the future. It’s still nowhere close to matching the instincts of human drivers, which is the single most important factor in road safety." For any model that makes predictions (be it an SVM, deep network, or 50ksloc of threshold functions written in C++), there are some set of parameters that one can use to tune the tradeoff between expected real-world true positives and false positives. SDC engineers have purposefully picked threshold values to favor conservative driving behavior and avoid crashes. See for yourself: In this video, the Waymo isn't avoiding a crash so much as potentially causing a rear-ender: https://youtu.be/vJovT6nsxB0?t=14 https://youtu.be/vJovT6nsxB0?t=14 I am no Waymo engineer, but the car is clearly mis-predicting the behavior of other threats in the intersection. Waymos tend to be "scared" because they're simply biased against being confident in their predictions about how clear the road is in front of them: https://www.youtube.com/watch?v=6itruDO57bk https://www.youtube.com/watch?v=6itruDO57bk Same with awkward lane changes. Waymos (and many other SDCs) have issues here because their engineers have convinced themselves via extensive manual testing and debugging that the cars would be less safe were they more "confident" in their predictions. The "villains" of self-driving know what's going on. The best we can do is to try to out-compete them; let them make us stronger. Passion for technology and passion for society are largely orthogonal. We want the winner to be somebody who gets both right.

cjhopman 8y ago
> In this video, the Waymo isn't avoiding a crash so much as potentially causing a rear-ender: https://youtu.be/vJovT6nsxB0?t=14 https://youtu.be/vJovT6nsxB0?t=14 wtf. This is a shitty example. If you either (1) read the captions on the video or (2) watch their full video where they talk about this, you'd realize that the waymo car stopped because a car off-camera ran the red into the intersection and that braking actually was avoiding a crash (or rather, decreasing the severity of a crash). https://youtu.be/spw176TZ7-8?t=41 https://youtu.be/spw176TZ7-8?t=41
- choppaface 8y ago
  I've watched this video several times. There is a red car clearly present in the intersection when the Waymo begins forward motion. That car partially occludes (and is a potential obstacle) for anything that's off to the left of the camera; it also is simply a dynamic obstacle in the intersection. The Waymo can certainly detect and predict if cross-traffic is going to roll through the intersection. But the Waymo started launching right when the light turned green... even though there's a car in the intersection at that time. I agree this is a bad example due to the camera angle, but there's enough evidence to show that the Waymo failed to correctly predict traffic here. And it executed a safety stop (which could have caused a rear-ender) as a result. Occluders are still really hard for SDCs. Humans can usually handle them well though. Handling occluders properly requires a very strong prediction system.

cjhopman 8y ago

> In this video, the Waymo isn't avoiding a crash so much as potentially causing a rear-ender: https://youtu.be/vJovT6nsxB0?t=14 https://youtu.be/vJovT6nsxB0?t=14 wtf. This is a shitty example. If you either (1) read the captions on the video or (2) watch their full video where they talk about this, you'd realize that the waymo car stopped because a car off-camera ran the red into the intersection and that braking actually was avoiding a crash (or rather, decreasing the severity of a crash). https://youtu.be/spw176TZ7-8?t=41 https://youtu.be/spw176TZ7-8?t=41

choppaface 8y ago
I've watched this video several times. There is a red car clearly present in the intersection when the Waymo begins forward motion. That car partially occludes (and is a potential obstacle) for anything that's off to the left of the camera; it also is simply a dynamic obstacle in the intersection. The Waymo can certainly detect and predict if cross-traffic is going to roll through the intersection. But the Waymo started launching right when the light turned green... even though there's a car in the intersection at that time. I agree this is a bad example due to the camera angle, but there's enough evidence to show that the Waymo failed to correctly predict traffic here. And it executed a safety stop (which could have caused a rear-ender) as a result. Occluders are still really hard for SDCs. Humans can usually handle them well though. Handling occluders properly requires a very strong prediction system.