-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
Profound progress towards FSD
- Thread starter Grifon
- Start date
diplomat33
Average guy who loves autonomous vehicles
I would add that even if the rewrite can't handle complex urban driving, if it just let us do the demo on autonomy day, hands-free, that would be really great IMO. I'd be happy.
mikes_fsd
Banned
My take on FSD is: If FSD only adds city driving to what we already get with AP or EAP, then for those of us that rarely go to a city, should we just resist the temptation to click the FSD $8k option? Or does the $8k option improve rural and interstate driving safety as well?
Something could be said for maintaining your city driving skills in case your FSD MS is in for repairs and you must drive an ICE car.
Something could be said for maintaining your city driving skills in case your FSD MS is in for repairs and you must drive an ICE car.
Last edited:
diplomat33
Average guy who loves autonomous vehicles
Had to grab a screenshot, just in case
I have never been against Tesla making progress towards FSD. And the autonomy day demo is a very simple, basic route. The roads were pretty simple and there was not a lot of traffic. I fully expect Tesla to be able to handle simple self-driving like that. Heck, they already demo'ed it on autonomy day so we know they can do it. Solving the more complex cases required for L5 autonomy will be a much greater challenge.
mikes_fsd
Banned
Can you please post each question in the right forum topic/thread? These have nothing to do with FSD or Autopilot.
diplomat33
Average guy who loves autonomous vehicles
I think by next year, FSD should improve rural and interstate driving quite a bit. So FSD should still be useful even if you don't do a lot of city driving. Moreover, as seen in the autonomy day demo, Tesla's intends to eventually make FSD be able to drive a route from start to finish, like from your house to whatever destination you choose. So FSD might still be useful for that as well.
diplomat33
Average guy who loves autonomous vehicles
The patent numbers are different. Maybe Tesla filed two different patents on the same invention or on very similar inventions?
Thanks. Looks like there's a related patent filed at the same time GENERATING GROUND TRUTH FOR MACHINE LEARNING FROM TIME SERIES ELEMENTS where the difference is on the last page of what is claimed.
willow_hiller
Well-Known Member
I would guess it's some limitation like "one concept per patent." So they wrote the legalese describing the 4D prediction and the video labeling in one patent, and then found they needed to file them as separate ideas.
diplomat33
Average guy who loves autonomous vehicles
Soda Popinski
Member
Patents aren't like journal articles. The important part is the Claims at the end. The rest is there to provide support for the claims, which are often argued back and forth between the inventor and the patent examiner. Patent attorneys often copy and paste the body from one related patent to the next for the same client, so it's very common to see what looks like the same patent, except for the abstract and claims. I don't think the examiners really care about the body of the patent, just the claims.
Tesla fans... its amazing. This is year 5. Five years. What makes this any more different than the 100+ other hyped updates. This is like the 100th hyped 'silky smooth' update. Seriously.
This whole 3D/4D hype is such nonsense. But ofcourse the tesla community lap it up.
This is nothing but this for labeling (16th min)
and something like vectornet for training.
Waypoint - The official Waymo blog: VectorNet: Predicting behavior to help the Waymo Driver make better decisions
Basically using splines aka vectors instead of raw pixels. Basically the spline has XYZ coordinates hence (3D) and then there's added time information using LSTM/RNN technique hence (4D).
But of-course Elon and co hypes it up as a holy grail as usual and the Tesla community is so gullible.
diplomat33
Average guy who loves autonomous vehicles
Thanks. In other words, Tesla is doing pretty standard machine learning but acting like it is some amazing radical breakthrough that will "solve FSD"?
willow_hiller
Well-Known Member
A good search engine is just natural language processing and a database. Anyone can do that, and there have been hundreds of academic papers written about it.
But if you do it at scale, then you've got Google. The constituent parts aren't revolutionary, but putting them together and doing it at scale is where the innovation happens.
powertoold
Active Member
Still to this day, the traffic control feature and cut-in detection are proofs of concept for FSD. I don't personally use the AP much since I enjoy driving my car, but I pay attention to the traffic control feature all the time, especially in confusing urban environments (with lights up and down hills and whatnot). It blows my mind that Tesla can release such a feature while Apple and Google still struggle with simple dictation (mostly Apple, lol).
The fleet is indispensable for FSD once you understand how they were able to develop the traffic control feature through fleet triggers and learning.
If you "read between the lines" from the traffic control and cut-in detection features, you'll see that a lot is possible once Tesla's FSD development stack is solidified with enough reliability and accuracy from fleet predictions.
The fleet is indispensable for FSD once you understand how they were able to develop the traffic control feature through fleet triggers and learning.
If you "read between the lines" from the traffic control and cut-in detection features, you'll see that a lot is possible once Tesla's FSD development stack is solidified with enough reliability and accuracy from fleet predictions.
Both Waymo with VectorNet and Tesla theoretically with the rewrite are using machine learning to predict trajectories just like both companies are using software. The details of how the networks are structured with inputs/outputs and sizes will be quite different, and even if those details were the same, the quality and quantity of training data will significantly affect the trained network.
Notably, it seems like Waymo simplifies lidar/sensor and static HD map data as inputs for the neural network to predict behaviors that are then passed on to the planner. Potentially with Tesla's approach, the neural network hidden layers are understanding the dynamic road layout and state, e.g., traffic lights and construction, so Autopilot could make a more reasonable prediction of a vehicle stopping and going at a temporary flashing red light. Yes, Waymo could also handle these situations by special casing behavior in the planner, but this is just one corner case and scaling it to handling thousands of other cases is probably why Tesla is trying to push more capabilities into the network.
powertoold
Active Member
If any of you do object oriented programming, you can better appreciate Tesla's approach to FSD. Essentially, what Tesla has been trying to accomplish is to create a perceptual prediction "object" using their strategies from operation vacation.
Once they're able to create this foundational "object" and validate its ability to deal with the long tail of events, they can create "subclasses" of this "object" to predict all manner static and moving features (object size, distance, and trajectories) related to driving.
When we were first introduced to this goal from Tesla at the autonomy day, Karpathy brought up the idea of cut in detection. When I saw that, I wasn't at all impressed because it seemed such a simple concept and feature. However, overtime I started to better understand the significance of what karpathy was saying at autonomy day. He talks a lot about using future data to label historical images. Tesla is essentially the only company right now able to have a consistent enough implementation of sensors to reliably use fleet data to train their FSD system. When most of us think about sensors related to FSD, we tend to only think about stuff like cameras and LiDAR and radar and ultrasonics, but we tend to neglect the other many sensors in the car, such as the sensors related to wheel position, pedal position, accelerometers, traction, barometers, light sensors, etc. When Tesla collects fleet data, they're able to utilize all sorts of measurements and sensors from the car, of which they have deep understanding and control over.
Once they're able to create this foundational "object" and validate its ability to deal with the long tail of events, they can create "subclasses" of this "object" to predict all manner static and moving features (object size, distance, and trajectories) related to driving.
When we were first introduced to this goal from Tesla at the autonomy day, Karpathy brought up the idea of cut in detection. When I saw that, I wasn't at all impressed because it seemed such a simple concept and feature. However, overtime I started to better understand the significance of what karpathy was saying at autonomy day. He talks a lot about using future data to label historical images. Tesla is essentially the only company right now able to have a consistent enough implementation of sensors to reliably use fleet data to train their FSD system. When most of us think about sensors related to FSD, we tend to only think about stuff like cameras and LiDAR and radar and ultrasonics, but we tend to neglect the other many sensors in the car, such as the sensors related to wheel position, pedal position, accelerometers, traction, barometers, light sensors, etc. When Tesla collects fleet data, they're able to utilize all sorts of measurements and sensors from the car, of which they have deep understanding and control over.
The current state of AP reminds me of a simple situation I faced a long time ago as a field tech. I brought this "top-of-the-line" fax modem to a client's home to replace his current one as he had connection problems. I plugged in the new modem and had issues with the handshake just like the old one. So I kept repeating how good this new modem was and all its capabilities etc, etc. He turned to me and said:"Sure, you keep telling me how good it is, but I see no proof of it". Simple, but right to the point. I kept quiet and eventually got it working reliably, but never forgot that simple observation. It really was one of the best on the market, but that meant zilch if it could not demonstrate that.
Right now, as far as I'm concerned, AP is a jack of some trades, and a master of none. I'd like it to be a jack of some trades and master of some. I'm not asking for perfection as that's unrealistic. Just some consistency. As I said elsewhere, from my experience, they've made good strides in the city. I've seen it stop for cars merging and avoided other situations on its own. It's also been less skittish around cars like in the past. All great and encouraging, however, I've also seen it completely neglect similar situations at other times. And the times when it did something, it was either too early or too late to react, or it overreacted. Thankfully, it's not often enough for me to give up on city driving. But on the freeway, it has gotten to the point that I no longer use it along certain stretches as phantom braking is guaranteed. And I also avoid driving close to larger vehicles as it will often brake, thinking they're cutting in,or at least, that's my best guess.
What also bothers me a little, is that with new updates, you'd think this behavior would improve, but it seemed to get worse. So, this re-write better be the savior. It looks great on paper and the demos are impressive, so I'm still hopeful.
Regardless of the eventual outcome though, this car is a blast to drive and the positives greatly outweighs any negatives. I'm willing to accept whatever becomes of FSD.
Right now, as far as I'm concerned, AP is a jack of some trades, and a master of none. I'd like it to be a jack of some trades and master of some. I'm not asking for perfection as that's unrealistic. Just some consistency. As I said elsewhere, from my experience, they've made good strides in the city. I've seen it stop for cars merging and avoided other situations on its own. It's also been less skittish around cars like in the past. All great and encouraging, however, I've also seen it completely neglect similar situations at other times. And the times when it did something, it was either too early or too late to react, or it overreacted. Thankfully, it's not often enough for me to give up on city driving. But on the freeway, it has gotten to the point that I no longer use it along certain stretches as phantom braking is guaranteed. And I also avoid driving close to larger vehicles as it will often brake, thinking they're cutting in,or at least, that's my best guess.
What also bothers me a little, is that with new updates, you'd think this behavior would improve, but it seemed to get worse. So, this re-write better be the savior. It looks great on paper and the demos are impressive, so I'm still hopeful.
Regardless of the eventual outcome though, this car is a blast to drive and the positives greatly outweighs any negatives. I'm willing to accept whatever becomes of FSD.
Similar threads
