3 ms·

I'm at probably 2,000 miles in AP on my Model 3, and at this point, I feel like I'm very knowledgeable about where I can trust AP and where I can't. I do fear

I'm at probably 2,000 miles in AP on my Model 3, and at this point, I feel like I'm very knowledgeable about where I can trust AP and where I can't. I do fear that people will give it too much trust in the beginning, and not try to learn how it behaves, because then, it's dangerous. Dying in a Tesla is relatively hard to do, but getting hurt or even just totaling your car because you didn't understand how AP works is bad (and dangerous for others, too!) I feel like maybe they shouldn't turn on AP until TACC has been active for some period of time, so that you can get used to the car and how it behaves. Even now, I am sometimes surprised (like the odd behavior when stopped traffic is ahead of you a ways, and the car right in front of you changes out of your line - your car doesn't notice the stopped traffic, so you have to stop yourself or you would slam into the stopped cars). I have to assume that, at some point, the front facing camera will supplement the radar and say, "hey, that looks like stopped traffic - I should stop", but I can understand, from a software development perspective, why that's a hard problem.

chli 8y ago
I don't own a Tesla but from what I understand you can get OTA updates which will change the way the AP is working. So if you need to build a mental model on how to "use" the AP to be safe and then the AP changes without notice I can see how that could end badly...
confiscate 8y ago
why would that be a hard problem? Shouldn't it be easy for the radar to detect if there are stopped obstacles ahead? I am not familiar with AP tech. So would appreciate if someone can explain why it's hard to detect stopped objects using radar.
- icebraining 8y ago
  Not an expert either, but the gist I got from previous threads is that the problem is distinguishing from objects which are obstacles from others which aren't (e.g. because you'll turn before hitting it). According to Techcrunch, that's also what happened in the Uber manslaughter - the system detected the pedestrian but didn't consider her a real obstacle to the car. That said, the impression I get is that Tesla's system is worse than even their initial version (provided by Mobileye).
  - carlmr 8y ago
    They need lidar/radar, then they have a chance to detect objects. Especially stopped traffic is a problem for vision only algorithms. It's easy to see moving objects, it's harder to see stopped objects. We humans are quite good at this, but machines have a long way to go here. Even with stereo view, everything that's a few hundred meters away doesn't have much difference between the left and right picture. Add to that that a few hundred meters away it's just a few pixels. So if it isn't moving it's like detecting a small object in a photo. Lidar/radar are the only sensors which can reliably detect stopped cars at a distance if you ask me.
    - zwily 8y ago
      They have always had radar. It’s not enough for detecting stopped vehicles at high speed though.
- drunken-serval 8y ago
  > Shouldn't it be easy for the radar to detect if there are stopped obstacles ahead? No. It's very difficult. Radar reflects off of everything. The only way to get a useful signal is to look for things that are moving at a different speed than you, like something moving relative to the road. So the radar software filters out all signals that have the same speed at the road ahead. This means the radar won't see anything stopped on the road, because it's filtering out a huge number of returns from the road. This gets even worse because Radar reflects very differently based off the material. It goes right through plastic, water (humans) absorb it. Even if you could try to read through the noise of the road, Radar doesn't have enough resolution to tell the difference between the road going uphill and a stopped firetruck.
  - confiscate 8y ago
    Thanks for the explanation. "radar won't see anything stopped on the road, because it's filtering out a huge number of returns from the road" what does that mean? sorry i am still confused. Isn't the whole point of radar to detect objects around you? It shouldn't matter if the object is in the road ahead, or on the side of the road, right? Because they are all obstacles that the car should avoid, no? Like, a stopped car in front of you, vs a tree next to you on the side of the road, are both things the software should avoid, and NOT filter out, right?
thrav 8y ago
I mean, I won’t claim it’s easy, but my Subaru can do it without breaking a sweat.

chli 8y ago

I don't own a Tesla but from what I understand you can get OTA updates which will change the way the AP is working. So if you need to build a mental model on how to "use" the AP to be safe and then the AP changes without notice I can see how that could end badly...

confiscate 8y ago

why would that be a hard problem? Shouldn't it be easy for the radar to detect if there are stopped obstacles ahead? I am not familiar with AP tech. So would appreciate if someone can explain why it's hard to detect stopped objects using radar.

icebraining 8y ago
Not an expert either, but the gist I got from previous threads is that the problem is distinguishing from objects which are obstacles from others which aren't (e.g. because you'll turn before hitting it). According to Techcrunch, that's also what happened in the Uber manslaughter - the system detected the pedestrian but didn't consider her a real obstacle to the car. That said, the impression I get is that Tesla's system is worse than even their initial version (provided by Mobileye).
- carlmr 8y ago
  They need lidar/radar, then they have a chance to detect objects. Especially stopped traffic is a problem for vision only algorithms. It's easy to see moving objects, it's harder to see stopped objects. We humans are quite good at this, but machines have a long way to go here. Even with stereo view, everything that's a few hundred meters away doesn't have much difference between the left and right picture. Add to that that a few hundred meters away it's just a few pixels. So if it isn't moving it's like detecting a small object in a photo. Lidar/radar are the only sensors which can reliably detect stopped cars at a distance if you ask me.
  - zwily 8y ago
    They have always had radar. It’s not enough for detecting stopped vehicles at high speed though.
drunken-serval 8y ago
> Shouldn't it be easy for the radar to detect if there are stopped obstacles ahead? No. It's very difficult. Radar reflects off of everything. The only way to get a useful signal is to look for things that are moving at a different speed than you, like something moving relative to the road. So the radar software filters out all signals that have the same speed at the road ahead. This means the radar won't see anything stopped on the road, because it's filtering out a huge number of returns from the road. This gets even worse because Radar reflects very differently based off the material. It goes right through plastic, water (humans) absorb it. Even if you could try to read through the noise of the road, Radar doesn't have enough resolution to tell the difference between the road going uphill and a stopped firetruck.
- confiscate 8y ago
  Thanks for the explanation. "radar won't see anything stopped on the road, because it's filtering out a huge number of returns from the road" what does that mean? sorry i am still confused. Isn't the whole point of radar to detect objects around you? It shouldn't matter if the object is in the road ahead, or on the side of the road, right? Because they are all obstacles that the car should avoid, no? Like, a stopped car in front of you, vs a tree next to you on the side of the road, are both things the software should avoid, and NOT filter out, right?

thrav 8y ago

I mean, I won’t claim it’s easy, but my Subaru can do it without breaking a sweat.