4 ms·

It is funny to me how most of the self-driving industry is hellbent on solving the long-tail problem while Tesla is stuck on…most of it. It’ll be interesting t

It is funny to me how most of the self-driving industry is hellbent on solving the long-tail problem while Tesla is stuck on…most of it. It’ll be interesting to see if they ever add LIDAR/re-add radar and admits their mistake.

XorNot 5y ago
Most of the self driving industry is not operating at scales like Tesla, or in as diverse environments. I'm extremely skeptical they're going to fair much better on scale up: its notable that the Uber crash had LIDAR and still hit a similar object (a metal mesh with lots of gaps and a person).
- ra7 5y ago
  Uber hit the person because they had disabled automatic emergency braking, not because the sensors failed to detect objects.
  - judge2020 5y ago
    Their Lidar wasn't plugged into AEB, so it's not like Lidar would have helped AEB either. The car's built in radar would have triggered a stop.
- andbberger 5y ago
  right, going to need more data to avoid running over a cyclist, as someone biking on the street is, as we all know, a black swan event. after all, who would want to bike somewhere with no protected bike infrastructure that allows shit like this to happen on public streets?
- upbeat_general 5y ago
  It seems that Tesla’s main issue is detection, not decision making (well at least their biggest issue). Sure, some environments like extreme weather are not being entered by most self-driving companies. However, their detection stacks in general work fine in the vast majority of places, it’s their decision making that needs work. As others pointed out, Uber failed not because of a lack of detection of the person. Moreover, the cases where Tesla publicly fails aren’t particularly challenging for detection (daylight, good weather).
  - XorNot 5y ago
    People treat this as though all Tesla's everywhere were for some reason never trained on a dataset including cyclists off to the side, and people make emotional statements of "obviously this is someone's fault for not dealing with this obvious scenario". But we all (should) know that's not how machine object detection systems work: they would reliably pass this sort of scenario all the time, and then some sort of adverse (input data wise) input situation from something humans don't even notice would trigger a complete failure. Which is to say it's not intuitively obvious that other systems won't have these sorts of failure scenarios - it certainly won't be proven till they have sufficiently widespread deployment to eek them out.
    - upbeat_general 5y ago
      Totally agree here. I do think redundant detection is important though and highlights the issue with the vision-only approach. If your object detection fails unexpectedly, as often happens with our current approaches, it's not catastrophic if you have another sensor modality to at least indicate there's a serious missed detection.
judge2020 5y ago
The cyclist was visible on the screen. This is obviously a case of "you don't get a good driving AI by hiring dozens of vision AI people".

XorNot 5y ago

Most of the self driving industry is not operating at scales like Tesla, or in as diverse environments. I'm extremely skeptical they're going to fair much better on scale up: its notable that the Uber crash had LIDAR and still hit a similar object (a metal mesh with lots of gaps and a person).

ra7 5y ago
Uber hit the person because they had disabled automatic emergency braking, not because the sensors failed to detect objects.
- judge2020 5y ago
  Their Lidar wasn't plugged into AEB, so it's not like Lidar would have helped AEB either. The car's built in radar would have triggered a stop.
andbberger 5y ago
right, going to need more data to avoid running over a cyclist, as someone biking on the street is, as we all know, a black swan event. after all, who would want to bike somewhere with no protected bike infrastructure that allows shit like this to happen on public streets?
upbeat_general 5y ago
It seems that Tesla’s main issue is detection, not decision making (well at least their biggest issue). Sure, some environments like extreme weather are not being entered by most self-driving companies. However, their detection stacks in general work fine in the vast majority of places, it’s their decision making that needs work. As others pointed out, Uber failed not because of a lack of detection of the person. Moreover, the cases where Tesla publicly fails aren’t particularly challenging for detection (daylight, good weather).
- XorNot 5y ago
  People treat this as though all Tesla's everywhere were for some reason never trained on a dataset including cyclists off to the side, and people make emotional statements of "obviously this is someone's fault for not dealing with this obvious scenario". But we all (should) know that's not how machine object detection systems work: they would reliably pass this sort of scenario all the time, and then some sort of adverse (input data wise) input situation from something humans don't even notice would trigger a complete failure. Which is to say it's not intuitively obvious that other systems won't have these sorts of failure scenarios - it certainly won't be proven till they have sufficiently widespread deployment to eek them out.
  - upbeat_general 5y ago
    Totally agree here. I do think redundant detection is important though and highlights the issue with the vision-only approach. If your object detection fails unexpectedly, as often happens with our current approaches, it's not catastrophic if you have another sensor modality to at least indicate there's a serious missed detection.

judge2020 5y ago

The cyclist was visible on the screen. This is obviously a case of "you don't get a good driving AI by hiring dozens of vision AI people".