Tesla to replace parking sensors with vision
Discussion
I’ll take the bait
I don’t believe radars are the future. Was driving in the rain today, based on the screen display the cameras were picking out the same cars in the same distance I could see. No concerns, no fantom braking.
I wonder how, in the future, additional infrared cameras could enable the car to see through fog or dense rain?
I don’t believe radars are the future. Was driving in the rain today, based on the screen display the cameras were picking out the same cars in the same distance I could see. No concerns, no fantom braking.
I wonder how, in the future, additional infrared cameras could enable the car to see through fog or dense rain?
It’s not vision OR radar, it’s Vision AND Radar. Everything vision can do, and then in the odd days where the environmental impact on vision degrades it’s performance, the radar may perform better,
It’s relatively easy to build a system for the happy day scenarios, it’s when things aren’t happy days that things get very tricky. What that means is you can build a vision only system that works maybe 99% of the time. That last 1% will stop you ever getting level 3 or above approval.
Oh, and the hor news is radar will be back come January. They’ve just filed papers in the US on the topic.
It’s relatively easy to build a system for the happy day scenarios, it’s when things aren’t happy days that things get very tricky. What that means is you can build a vision only system that works maybe 99% of the time. That last 1% will stop you ever getting level 3 or above approval.
Oh, and the hor news is radar will be back come January. They’ve just filed papers in the US on the topic.
Heres Johnny said:
Oh, and the hor news is radar will be back come January. They’ve just filed papers in the US on the topic.
Sounds exciting!Although digging a bit, it might not be as simple as “radar will be back” as their admission Tesla were wrong to rely on just cameras.
Sounds more like an extension to a confidential agreement. It appears their “new” radar was submitted 6 months ago before they disabled the current radar and decided to rely solely on cameras. If it requires a confidential agreement, excited to see what this new system can do.
If it’s a new radar, I’m guessing it won’t be retrospectively fitted to our cars. I’ll make a guess as to the strategy behind this;
- Radar only to their cheaper new model, FSD won’t be available.
https://insideevs.com/news/625502/tesla-alerts-fcc...
Edited by Hans_Gruber on Thursday 8th December 06:51
z4RRSchris said:
front distance is a guess at best, i parked up and it said 14 inches but in reality was about 2.
Yeah, that RSymons video shows plenty of examples where the indicated distance is a lot larger than the actual, measured clearance. So if you can't rely on the numbers that it's showing then it's useless IMO. mids said:
z4RRSchris said:
front distance is a guess at best, i parked up and it said 14 inches but in reality was about 2.
Yeah, that RSymons video shows plenty of examples where the indicated distance is a lot larger than the actual, measured clearance. So if you can't rely on the numbers that it's showing then it's useless IMO. It will only worsen their already damaged reputation.
Tesla Vision parking sensors highlghts the whole problem of AI in this space.
With USS we kind of know the limitations. We know where the sensors are and if we drove directly a post in the centre of the car it would be less accurate than if the post was in front of the sensor, we know they're located 50cm in the air, so a low curb you might be more mindful of.. it's all pretty predictable.
With AI models you have no idea what it's thinking or why, just the end result. You don't know what combination of circumstanes it's better or worse at, you've an idea that if its pitch dark or pouring with rain it's going to struggle, but what type of bollard is it good and bad at? Does the relative distance, colour, height, or anything else that not what you're trying to avoid have a baring on the performance? And then to top it all, every time they tweak the thing by putting a bit more data through the training model, all the rules will change slightly. It is the nature of AI.
If you think the reliability of parking sensors is a challenge for Tesla to demonstrate, the whole driving thing is an even bigger one. Every change could have a negative consequence elsewhere, and because its not programatic and rules based, you've absolutely no idea where that might be. Symons certainly hasn't a clue on any of this.
With USS we kind of know the limitations. We know where the sensors are and if we drove directly a post in the centre of the car it would be less accurate than if the post was in front of the sensor, we know they're located 50cm in the air, so a low curb you might be more mindful of.. it's all pretty predictable.
With AI models you have no idea what it's thinking or why, just the end result. You don't know what combination of circumstanes it's better or worse at, you've an idea that if its pitch dark or pouring with rain it's going to struggle, but what type of bollard is it good and bad at? Does the relative distance, colour, height, or anything else that not what you're trying to avoid have a baring on the performance? And then to top it all, every time they tweak the thing by putting a bit more data through the training model, all the rules will change slightly. It is the nature of AI.
If you think the reliability of parking sensors is a challenge for Tesla to demonstrate, the whole driving thing is an even bigger one. Every change could have a negative consequence elsewhere, and because its not programatic and rules based, you've absolutely no idea where that might be. Symons certainly hasn't a clue on any of this.
Gassing Station | Tesla | Top of Page | What's New | My Stuff