4 ms·

It's kind of interesting that this appears to happen with police cars and emergency vehicles mostly. Is there a particular way of parking such vehicles that mak

It's kind of interesting that this appears to happen with police cars and emergency vehicles mostly. Is there a particular way of parking such vehicles that makes it hard(er) for a Tesla to detect it? It might also indeed be about phantom braking. My relatively old (probably designed around 10 to 20 years ago) system gets confused sometimes when I pass a truck or a bus and once even when I passed a bicycle. It happened at least 5 times in less than 10,000km. So I suspect there is some merit to the Autopilot disabling the feature.

Rebelgecko 5y ago
I wonder if the cameras or ML are thrown off by flashing emergency lights or emergency vehicles' livery (most cars on the road use just 1 paint color)
- ajross 5y ago
  Yeah. I'm as big a Tesla booster as you'll find but I'm in the "this is probably a bug in recognition" camp. They can presumably fix this via properly parametrized modelling. It's not a rare edge case at all.
  - myself248 5y ago
    I have a hunch that at some point, there'll be an investigation and a discovery process, and we'll learn that the training dataset deliberately _excluded_ emergency vehicles because they didn't represent "normal driving" or something. That's purely uninformed speculation, but I could see it happening like this: Early on, someone says "The totality of situations encountered on the road is too complex. We're not building a system to do everything, we're building it to take over during the boring parts of a drive. Only teach it the boring stuff." Then a few years later, someone else, forgetting those initial assumptions, says "Hey this thing has done like a billion miles of boring stuff, why wouldn't you trust it for more?"
    - ajross 5y ago
      No, the car sees emergency vehicles. It sees stopped vehicles. It sees vehicles stopped across lanes. This situation is far, far more common than the count of actual accidents we've seen. There probably is a bug, but it's not remotely as simple as "can't see fire engines".
  - antattack 5y ago
    Until recently Tesla relied on forward microwave radar to detect collisions which is very bad at detecting stationary objects. Great majority of stationary cars on a highway are emergency vehicles, hence the 14 or so collisions.
andrewia 5y ago
I think the flashing lights and appearance might degrade autopilot. Other cars also have issues because they use radar (except for some new models with camera-based systems like the 2022 Honda Civic) and ignore stationary targets when they see a moving vehicle as a more likely "target". Some systems use a combination of radar and camera (with camera providing low-speed, parked car, bike, and pedestrian detection). Then it's up to the quality of the algorithms and tests have shown that some cars are pretty poor at detection.
- mindslight 5y ago
  The human visual system is amazing at normalizing across many orders of magnitudes (cf HDR imaging), but yet with the evolution of LED lights and a lack of accountability, the lights on police cars have gotten way too bright for even human drivers at night. I wouldn't be surprised if the ultimate cause is a complete swamping of the cameras for everything that isn't immediately in front of the car, and even the lane lines disappearing when close enough to the blinding lights. What is a computer supposed to do when confronted with complete loss of visual input? A human will squint, slow down by 10-20mph, extrapolate the lane lines, and play the odds. But there is no surefire way to proceed safely, and thus no straightforward action to program a computer with.
  - choeger 5y ago
    Well, if the computer can recognize the situation it can always slow down, alert the driver, and eventually stop. That's the nice thing about cars: There's always a safe default mode. I suspect that the computer cannot recognize this problem reliably, though.
    - mindslight 5y ago
      Slowing down to under 40mph is not a safe thing to do on a highway and is illegal in many places. And suddenly slowing down is extra unsafe when there's a car right behind you that's also being blinded.
      - 5y ago

Rebelgecko 5y ago

I wonder if the cameras or ML are thrown off by flashing emergency lights or emergency vehicles' livery (most cars on the road use just 1 paint color)

ajross 5y ago
Yeah. I'm as big a Tesla booster as you'll find but I'm in the "this is probably a bug in recognition" camp. They can presumably fix this via properly parametrized modelling. It's not a rare edge case at all.
- myself248 5y ago
  I have a hunch that at some point, there'll be an investigation and a discovery process, and we'll learn that the training dataset deliberately _excluded_ emergency vehicles because they didn't represent "normal driving" or something. That's purely uninformed speculation, but I could see it happening like this: Early on, someone says "The totality of situations encountered on the road is too complex. We're not building a system to do everything, we're building it to take over during the boring parts of a drive. Only teach it the boring stuff." Then a few years later, someone else, forgetting those initial assumptions, says "Hey this thing has done like a billion miles of boring stuff, why wouldn't you trust it for more?"
  - ajross 5y ago
    No, the car sees emergency vehicles. It sees stopped vehicles. It sees vehicles stopped across lanes. This situation is far, far more common than the count of actual accidents we've seen. There probably is a bug, but it's not remotely as simple as "can't see fire engines".
- antattack 5y ago
  Until recently Tesla relied on forward microwave radar to detect collisions which is very bad at detecting stationary objects. Great majority of stationary cars on a highway are emergency vehicles, hence the 14 or so collisions.

andrewia 5y ago

I think the flashing lights and appearance might degrade autopilot. Other cars also have issues because they use radar (except for some new models with camera-based systems like the 2022 Honda Civic) and ignore stationary targets when they see a moving vehicle as a more likely "target". Some systems use a combination of radar and camera (with camera providing low-speed, parked car, bike, and pedestrian detection). Then it's up to the quality of the algorithms and tests have shown that some cars are pretty poor at detection.

mindslight 5y ago
The human visual system is amazing at normalizing across many orders of magnitudes (cf HDR imaging), but yet with the evolution of LED lights and a lack of accountability, the lights on police cars have gotten way too bright for even human drivers at night. I wouldn't be surprised if the ultimate cause is a complete swamping of the cameras for everything that isn't immediately in front of the car, and even the lane lines disappearing when close enough to the blinding lights. What is a computer supposed to do when confronted with complete loss of visual input? A human will squint, slow down by 10-20mph, extrapolate the lane lines, and play the odds. But there is no surefire way to proceed safely, and thus no straightforward action to program a computer with.
- choeger 5y ago
  Well, if the computer can recognize the situation it can always slow down, alert the driver, and eventually stop. That's the nice thing about cars: There's always a safe default mode. I suspect that the computer cannot recognize this problem reliably, though.
  - mindslight 5y ago
    Slowing down to under 40mph is not a safe thing to do on a highway and is illegal in many places. And suddenly slowing down is extra unsafe when there's a car right behind you that's also being blinded.
    - 5y ago