carsly
Well-known member
- First Name
- Vin
- Joined
- Dec 13, 2023
- Threads
- 93
- Messages
- 1,529
- Reaction score
- 2,898
- Location
- Princeton, NJ
- Vehicles
- LR Defender, CT AWD
- Thread starter
- #1
As we await the launch of unsupervised FSD (at least in Austin for Cybercab) in a few weeks I've been spending some time digging in at the edges of Tesla's software to see how it can interpret our unstructured world, especially in non-paved, non-marked roads and I stumbled on an example right in my own driveway.
Situation, i have a circular driveway with a straight section on one side. Everything except where I park in front of the garage is gravel. No markings, no lines, just gravel. We've also got a lot of trees, grass and shrubbery around and as I was reversing in this morning the parking animation caught my eye. It's "seeing" quite a bit and interpreting in real-time. Trees (including defining the size and location of trunks), shrubs, the edge of the gravel, the width of the gravel, etc. are all clearly captured and interpreted. While this make seem obvious in a paved parking lot with painted lines, this situation is the exact opposite.
So if we've got this high-resolution mapping in 3D at low speeds, how clearly is it seeing at road speeds? We have the overly simplified FSD visualization, but my gut tells me its still seeing and interpreting something close to this high-resolution parking view in real-time - which is crazy not only because it has to map and interpret but also determine actions - accelerator, brake, turning input, etc. in real-time as well. Here is where efficient, and accurate, processing matters. We can debate whether camera-only is the right approach due to occlusions, reflections, rain, etc. but simplifying the models to one time of input - visual data - against a neural net for interpretation sounded batsh*t crazy but might be the only really efficient and scalable way to solve autonomous driving.
PS, I really liked the radar in my HW3 Model S.
Situation, i have a circular driveway with a straight section on one side. Everything except where I park in front of the garage is gravel. No markings, no lines, just gravel. We've also got a lot of trees, grass and shrubbery around and as I was reversing in this morning the parking animation caught my eye. It's "seeing" quite a bit and interpreting in real-time. Trees (including defining the size and location of trunks), shrubs, the edge of the gravel, the width of the gravel, etc. are all clearly captured and interpreted. While this make seem obvious in a paved parking lot with painted lines, this situation is the exact opposite.
So if we've got this high-resolution mapping in 3D at low speeds, how clearly is it seeing at road speeds? We have the overly simplified FSD visualization, but my gut tells me its still seeing and interpreting something close to this high-resolution parking view in real-time - which is crazy not only because it has to map and interpret but also determine actions - accelerator, brake, turning input, etc. in real-time as well. Here is where efficient, and accurate, processing matters. We can debate whether camera-only is the right approach due to occlusions, reflections, rain, etc. but simplifying the models to one time of input - visual data - against a neural net for interpretation sounded batsh*t crazy but might be the only really efficient and scalable way to solve autonomous driving.
PS, I really liked the radar in my HW3 Model S.
Sponsored
