4 ms·

The article linked mentioned "B was the problem", meaning it ignored the sensor input. Another article(1) about it mentions that Uber had reduced the number of

The article linked mentioned "B was the problem", meaning it ignored the sensor input. Another article(1) about it mentions that Uber had reduced the number of LIDAR sensors from five to just one rooftop sensor. The president of the company that builds the sensors even said you would need side sensors to see pedestrians, especially at night. But LIDAR is expensive. So it seems that, like Tesla, they were trying to rely on radar for that kind of thing. But that does leave the blind spot(2) I mentioned. Given that I don't see how this case is so different. Though I still don't understand how, specifically, that blind spot problem works. They say when the car in front of you changes lanes and a stopped truck is now in your lane you'll hit it. What if no car changed lanes in front of you? Would you still hit a stopped truck in your lane? The guy whose Tesla hit the bridge column seems to tell me yes, but I'm not sure. 1. https://www.theregister.co.uk/2018/03/28/uber_selfdriving_death_may_have_been_due_to_lidar_blind_spot/ https://www.theregister.co.uk/2018/03/28/uber_selfdriving_de... 2. https://www.wired.com/story/tesla-autopilot-why-crash-radar/ https://www.wired.com/story/tesla-autopilot-why-crash-radar/

stevenwoo 8y ago
I thought people were saying in previous comment threads that Teslas ignore stationary objects because too many false positives from roadside objects that are not in the road. The combination of software and sensors on Teslas is not enough to be able to distinguish stationary objects in the path versus stationary objects not in the path. Based on the slim evidence of the Chinese accident video and the LA post accident photo, it looks like that's correct - http://www.dailymotion.com/video/x4tes8u http://www.dailymotion.com/video/x4tes8u https://cdn.teslarati.com/wp-content/uploads/2018/01/tesla-crash-fire-truck-405-los-angeles.jpg https://cdn.teslarati.com/wp-content/uploads/2018/01/tesla-c...
- wilburTheDog 8y ago
  Wow. That Chinese accident video confirms my worst possible understanding of this. There are no special conditions. If a truck is parked in your lane you're going to hit it. More people should know about this. Maybe people don't want to believe it. I don't know. But I do know that I'm not going to trust any system that doesn't have multiple LIDAR sensors. Though it may take a while for the price on those to drop enough to make it affordable.
Symmetry 8y ago
I haven't worked with automotive radars but the way it works in aeronautical radar is that you apply frequency binning in the electronics and ignore returns that have the wrong Doppler shift at the waveform level, long before step A in the article. If you didn't do it this way your computer would be overwhelmed considering every tree and house you can see but only moving objects like airplanes. It's the same principle as the way your eyes only report things that move over your retina as your eye moves, letting you filter out imperfections in the lens of your eye and your blind spot. Similarly, I assume that the computers in a car would be overwhelmed by the task of taking the returns from every single post on a guardrail and sorting them into distinct objects before deciding to ignore them due to relative velocity so I'm fairly sure that a similar mechanism is at work within the hardware of the radar unit the Uber car. So given that the woman showed up on Uber's sensors she would have been detected by the camera system or the lidar. And apparently she was detected.
- wilburTheDog 8y ago
  Thanks for the explanation. I hadn't heard of frequency binning(1) before. It makes sense that stationary objects would be filtered out before they even reach the signal processing software. But Uber did seem to be relying more heavily on radar than the other sensors. From an article I linked in another comment "the number of LIDAR sensors were reduced from five to just one – mounted on the roof – and in their place, the number of radar sensors was increased from seven to 10. Uber also reduced the number of cameras on the car from 20 to seven." And for LIDAR this "results in a blind spot low to the ground all around the car." They had to know radar wouldn't detect stationary objects. So the signal processor should be prioritizing camera and LIDAR reports of anything stationary in the road. If it really was programmed to ignore such a signal it seems like gross malfeasance to me. If not, then the speculation in the linked article is incorrect and it's the same problem Tesla's system has. 1. https://www.eetimes.com/document.asp?doc_id=1278779 https://www.eetimes.com/document.asp?doc_id=1278779

stevenwoo 8y ago

I thought people were saying in previous comment threads that Teslas ignore stationary objects because too many false positives from roadside objects that are not in the road. The combination of software and sensors on Teslas is not enough to be able to distinguish stationary objects in the path versus stationary objects not in the path. Based on the slim evidence of the Chinese accident video and the LA post accident photo, it looks like that's correct - http://www.dailymotion.com/video/x4tes8u http://www.dailymotion.com/video/x4tes8u https://cdn.teslarati.com/wp-content/uploads/2018/01/tesla-crash-fire-truck-405-los-angeles.jpg https://cdn.teslarati.com/wp-content/uploads/2018/01/tesla-c...

wilburTheDog 8y ago
Wow. That Chinese accident video confirms my worst possible understanding of this. There are no special conditions. If a truck is parked in your lane you're going to hit it. More people should know about this. Maybe people don't want to believe it. I don't know. But I do know that I'm not going to trust any system that doesn't have multiple LIDAR sensors. Though it may take a while for the price on those to drop enough to make it affordable.

Symmetry 8y ago

I haven't worked with automotive radars but the way it works in aeronautical radar is that you apply frequency binning in the electronics and ignore returns that have the wrong Doppler shift at the waveform level, long before step A in the article. If you didn't do it this way your computer would be overwhelmed considering every tree and house you can see but only moving objects like airplanes. It's the same principle as the way your eyes only report things that move over your retina as your eye moves, letting you filter out imperfections in the lens of your eye and your blind spot. Similarly, I assume that the computers in a car would be overwhelmed by the task of taking the returns from every single post on a guardrail and sorting them into distinct objects before deciding to ignore them due to relative velocity so I'm fairly sure that a similar mechanism is at work within the hardware of the radar unit the Uber car. So given that the woman showed up on Uber's sensors she would have been detected by the camera system or the lidar. And apparently she was detected.

wilburTheDog 8y ago
Thanks for the explanation. I hadn't heard of frequency binning(1) before. It makes sense that stationary objects would be filtered out before they even reach the signal processing software. But Uber did seem to be relying more heavily on radar than the other sensors. From an article I linked in another comment "the number of LIDAR sensors were reduced from five to just one – mounted on the roof – and in their place, the number of radar sensors was increased from seven to 10. Uber also reduced the number of cameras on the car from 20 to seven." And for LIDAR this "results in a blind spot low to the ground all around the car." They had to know radar wouldn't detect stationary objects. So the signal processor should be prioritizing camera and LIDAR reports of anything stationary in the road. If it really was programmed to ignore such a signal it seems like gross malfeasance to me. If not, then the speculation in the linked article is incorrect and it's the same problem Tesla's system has. 1. https://www.eetimes.com/document.asp?doc_id=1278779 https://www.eetimes.com/document.asp?doc_id=1278779