7 ms·

I want to ask: why is Tesla's advanced suite of cameras and ultrasonic sensors unable to detect a big, yellow and black stationary barrier? In combination with

by dannyw 9y ago

I want to ask: why is Tesla's advanced suite of cameras and ultrasonic sensors unable to detect a big, yellow and black stationary barrier? In combination with chevron markings that mean "never ever drive here". Either should be a globally overriding signal to emergency break or turn, not "keep following the white line". If the answer is "during morning light, it doesn't see it", then they need to take the feature offline until the car is fitted with an appropriate LIDAR; or other technology that can detect stationary obstacles.

jonknee 9y ago
I've seen a lot of Tesla apologizers talk about how it's such a bad design, but people seem to be able to dodge it successfully and so should Autopilot. It seems like this feature should require hands on the wheel until it has enough training data that it doesn't do things like smash you into an obvious obstacle in whatever lighting conditions are possible at any time.
- pkulak 9y ago
  > this feature should require hands on the wheel It does.
  - jonknee 9y ago
    Yet, in this fatal crash: > "the driver’s hands were not detected on the wheel for six seconds prior to the collision" https://www.tesla.com/blog/update-last-week%E2%80%99s-accident https://www.tesla.com/blog/update-last-week%E2%80%99s-accide...
    - ricardobeat 9y ago
      What is the car supposed to do? Have robot arms that force your hands? It will actually slow down and stop if you continue not holding the wheel.
- rasz 9y ago
  > but people seem to be able to dodge it successfully are they? so why was the barrier crushed in the first place before the crash? and why was it designed to be easily rebuild? its almost like someone knew people will have trouble and will eventually drive into it.
unityByFreedom 9y ago
Apparently their radar can only see objects that move. So, to see this barrier, you'd have to guess its distance from 2D images. That's a hard problem. I think this would be called "obstacle detection using stereo vision" or maybe even "monocular vision" if they're using one camera. There isn't as much research into this field as there is in systems that use lidar because lidar has been proven to work, and academics probably don't care as much about building their products at massive scales, so the cost of lidar doesn't matter as much. Searching Arxiv for that phrase [1], the top result is a paper called A Joint 3D-2D based Method for Free Space Detection on Roads [2] that uses Lidar. The next result that doesn't mention lidar is Free Space Estimation using Occupancy Grids and Dynamic Object Detection [3]. Following that is Failing to Learn: Autonomously Identifying Perception Failures for Self-driving Cars [4]. In that paper's conclusion, it states its next plan is to use lidar to improve data quality, > we plan to incorporate free space computation from the path of the vehicle and from active sensor returns like LIDAR to identify false positives to further improve our assessment and understanding of modern object detectors at the fleet level. [1] https://arxiv.org/find/all/1/all:+AND+vision+AND+stereo+AND+object+detection/0/1/0/all/0/1 https://arxiv.org/find/all/1/all:+AND+vision+AND+stereo+AND+... [2] https://arxiv.org/pdf/1711.02144.pdf https://arxiv.org/pdf/1711.02144.pdf [3] https://arxiv.org/pdf/1708.04989.pdf https://arxiv.org/pdf/1708.04989.pdf [4] https://arxiv.org/pdf/1707.00051.pdf https://arxiv.org/pdf/1707.00051.pdf
YeGoblynQueenne 9y ago
>> I want to ask: why is Tesla's advanced suite of cameras and ultrasonic sensors unable to detect a big, yellow and black stationary barrier? The answer to that is that sensors detect, but don't make decisions. The decisions are made by the car AI that takes input from the censors and outputs control signals. So the pertinent answer is "why is Tesla's AI unable to make the decisions necessary to avoid a big, yellow and black stationary barrier?". Unfortunately, nobody can answer that, including Tesla engineers, because their cars' AI is trained with black-box machine learning algorithms that are extremely resistant to any attempt to understand how they make decisions. But, you know. I'm pretty sure they're really safe /s
- danso 9y ago
  Tesla's sensors are also claimed to detect/read, but not recognize or react to imagery of stop signs and traffic lights: https://forums.tesla.com/forum/forums/autopilot-handling-stop-signs-and-traffic-signals https://forums.tesla.com/forum/forums/autopilot-handling-sto...
- protomyth 9y ago
  > Unfortunately, nobody can answer that, including Tesla engineers, because their cars' AI is trained with black-box machine learning algorithms that are extremely resistant to any attempt to understand how they make decisions. What is the justification for allowing anything we don't understand to be on the road? I really want the NTSB to ask specifically what decisions the Tesla was making and not accept "black box learning" as an answer.
  - dragonwriter 9y ago
    > What is the justification for allowing anything we don't understand to be on the road? What's the justification for allowing human brains in the control loop?
    - protomyth 9y ago
      Because we have a right to travel in most countries and cars are the mechanism that we use and we can be held responsible for our actions. This black box AI has none of those attributes.
rasz 9y ago
Did you mean small, dirty crushed piece of metal and invisible hashed out chevrons? Btw you wont find this type of barrier in this state of neglect in EU, its illegal, its like setting up a trap to kill people.

by dannyw 9y ago

jonknee 9y ago

I've seen a lot of Tesla apologizers talk about how it's such a bad design, but people seem to be able to dodge it successfully and so should Autopilot. It seems like this feature should require hands on the wheel until it has enough training data that it doesn't do things like smash you into an obvious obstacle in whatever lighting conditions are possible at any time.

pkulak 9y ago
> this feature should require hands on the wheel It does.
- jonknee 9y ago
  Yet, in this fatal crash: > "the driver’s hands were not detected on the wheel for six seconds prior to the collision" https://www.tesla.com/blog/update-last-week%E2%80%99s-accident https://www.tesla.com/blog/update-last-week%E2%80%99s-accide...
  - ricardobeat 9y ago
    What is the car supposed to do? Have robot arms that force your hands? It will actually slow down and stop if you continue not holding the wheel.
rasz 9y ago
> but people seem to be able to dodge it successfully are they? so why was the barrier crushed in the first place before the crash? and why was it designed to be easily rebuild? its almost like someone knew people will have trouble and will eventually drive into it.

unityByFreedom 9y ago

Apparently their radar can only see objects that move. So, to see this barrier, you'd have to guess its distance from 2D images. That's a hard problem. I think this would be called "obstacle detection using stereo vision" or maybe even "monocular vision" if they're using one camera. There isn't as much research into this field as there is in systems that use lidar because lidar has been proven to work, and academics probably don't care as much about building their products at massive scales, so the cost of lidar doesn't matter as much. Searching Arxiv for that phrase [1], the top result is a paper called A Joint 3D-2D based Method for Free Space Detection on Roads [2] that uses Lidar. The next result that doesn't mention lidar is Free Space Estimation using Occupancy Grids and Dynamic Object Detection [3]. Following that is Failing to Learn: Autonomously Identifying Perception Failures for Self-driving Cars [4]. In that paper's conclusion, it states its next plan is to use lidar to improve data quality, > we plan to incorporate free space computation from the path of the vehicle and from active sensor returns like LIDAR to identify false positives to further improve our assessment and understanding of modern object detectors at the fleet level. [1] https://arxiv.org/find/all/1/all:+AND+vision+AND+stereo+AND+object+detection/0/1/0/all/0/1 https://arxiv.org/find/all/1/all:+AND+vision+AND+stereo+AND+... [2] https://arxiv.org/pdf/1711.02144.pdf https://arxiv.org/pdf/1711.02144.pdf [3] https://arxiv.org/pdf/1708.04989.pdf https://arxiv.org/pdf/1708.04989.pdf [4] https://arxiv.org/pdf/1707.00051.pdf https://arxiv.org/pdf/1707.00051.pdf

YeGoblynQueenne 9y ago

>> I want to ask: why is Tesla's advanced suite of cameras and ultrasonic sensors unable to detect a big, yellow and black stationary barrier? The answer to that is that sensors detect, but don't make decisions. The decisions are made by the car AI that takes input from the censors and outputs control signals. So the pertinent answer is "why is Tesla's AI unable to make the decisions necessary to avoid a big, yellow and black stationary barrier?". Unfortunately, nobody can answer that, including Tesla engineers, because their cars' AI is trained with black-box machine learning algorithms that are extremely resistant to any attempt to understand how they make decisions. But, you know. I'm pretty sure they're really safe /s

danso 9y ago
Tesla's sensors are also claimed to detect/read, but not recognize or react to imagery of stop signs and traffic lights: https://forums.tesla.com/forum/forums/autopilot-handling-stop-signs-and-traffic-signals https://forums.tesla.com/forum/forums/autopilot-handling-sto...
protomyth 9y ago
> Unfortunately, nobody can answer that, including Tesla engineers, because their cars' AI is trained with black-box machine learning algorithms that are extremely resistant to any attempt to understand how they make decisions. What is the justification for allowing anything we don't understand to be on the road? I really want the NTSB to ask specifically what decisions the Tesla was making and not accept "black box learning" as an answer.
- dragonwriter 9y ago
  > What is the justification for allowing anything we don't understand to be on the road? What's the justification for allowing human brains in the control loop?
  - protomyth 9y ago
    Because we have a right to travel in most countries and cars are the mechanism that we use and we can be held responsible for our actions. This black box AI has none of those attributes.

rasz 9y ago

Did you mean small, dirty crushed piece of metal and invisible hashed out chevrons? Btw you wont find this type of barrier in this state of neglect in EU, its illegal, its like setting up a trap to kill people.