-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
Elon: FSD Beta tweets
- Thread starter Terminator857
- Start date
I see it as easy pickings. There's no wiggle room over stopping at stop signs in law so that was a slam dunk for the NHTSA. Plus they picked on several other of what they perceived to be FSD's flaws.
Why didn't they care about risky behavior at roundabouts or passing school buses or whatever? I kind of thought they would, issue some further requests for performance improvement on a regular basis, but they didn't follow-up. That's what's curious to me, that they didn't continue to press Tesla for change.
I'm not even sure we got a public acknowledgment that the NHTSA agreed that Tesla has fixed everything that was in the recall, and that the bad behaviors will not reemerge.
NHTSA stepped in because Tesla's policy was to roll stop signs. So NHTSA is not jumping on the fact that FSDb can't do a good job somewhere, but that Tesla's policy was counter to law.
There are currently no known L3+ AVs on the road that are travelling faster than 65MPH, so we don't know how the big boys (Waymo, Cruise, Zoox, and even the latest L3 Mercedes) will handle freeway speeds. My guess is that they will be required to adhere to posted speed limits, especially freeway speed limits that have "MAX" in the sign. I don't imagine any government regulatory body allowing AVs while in L3+ to exceed the speed limit on their own.
People can argue all they want that driving faster than the speed limit is "safe" because the flow of traffic is moving faster, but unless traffic laws change to address this, the law trumps perceived safety. Case in point: I encourage you all to flag down an officer (I usually speak to them while they're at a gas station, since they're not really busy at that moment) and ask them "Is driving over the speed limit okay if the flow of traffic is doing it?" In SoCal, every officer I've asked that question has said "No". I've even called my local highway patrol office and asked the desk clerk the question - same answer: No.
This is usually met with "But of course police are going to say it's illegal, but that doesn't change the fact that it's dangerous to drive the speed limit when everyone else is going faster." To which I will direct you NHTSA and your state highway patrol to ask them the question.
All that said, unless laws change in the near future, Tesla will drive the speed limit when they release L3. In the meantime, in L2, where the driver is still in control, you can set the speed, as you're the one breaking the law, and responsible for the consequences. This new drive with the flow of traffic feature may be short lived if it's released to the wild with no way to disable it or override the behavior.
People can argue all they want that driving faster than the speed limit is "safe" because the flow of traffic is moving faster, but unless traffic laws change to address this, the law trumps perceived safety. Case in point: I encourage you all to flag down an officer (I usually speak to them while they're at a gas station, since they're not really busy at that moment) and ask them "Is driving over the speed limit okay if the flow of traffic is doing it?" In SoCal, every officer I've asked that question has said "No". I've even called my local highway patrol office and asked the desk clerk the question - same answer: No.
This is usually met with "But of course police are going to say it's illegal, but that doesn't change the fact that it's dangerous to drive the speed limit when everyone else is going faster." To which I will direct you NHTSA and your state highway patrol to ask them the question.
All that said, unless laws change in the near future, Tesla will drive the speed limit when they release L3. In the meantime, in L2, where the driver is still in control, you can set the speed, as you're the one breaking the law, and responsible for the consequences. This new drive with the flow of traffic feature may be short lived if it's released to the wild with no way to disable it or override the behavior.
In theory this entirely training-data driven system should be able to adapt to different driving locations, or ideally train on the entire world. Location specific driving customs and road layouts can be thought of as two dimensions of a vast multidimensional array. Other environmental dimensions would be time of day, weather, traffic density, construction and so on. But these conditions, even though they may persist for actimespan of minutes to hours, are also not really independent of the moment-by-moment events and driving decisions that must go into the solution.
Everything interacts with everything, so there's really no special gotcha associated with a different location, even though it's true that laws might be different. Mathematically, we might say that the perimeters of location, laws or environment conditions are not orthogonal i e. not independently varying vs. one another. Nor are they orthogonal to any other aspect of the driving task.
True, the model could be smaller if we remove some of these variations, and restricting the training data geographically is one way to do that. But it's not obvious to me that the compute or memory savings would be a key enabler of FSD.
I think at this time, the bigger reason for a geographical restriction is not so much for model simplification, but rather the simple fact that Tesla has not been running FSD on a large scale outside of North America, so the available training data set is much smaller elsewhere.
But let's say the data is available worldwide and think about whether a unified world driving model is achievable. Within the USA alone, there are enough differences both legal and empirical that it would need training scenarios to cover the gamut. What we outsiders don't know yet is whether the machine language model, running on the computer in the car, would be sufficient to cover scenarios from all over the USA much less the entire world.. And further, if worldwide capability is too much for the HW3 computer, then is it really true that resorting to customized regional software releases would solve that problem?
If so, great and it's no big disadvantage but it's a little less elegant and complicates both the training and the software version rollouts & maintenance.
By the same token, if the USA and Canada have enough regional differences to require software package downloads as one moves across the country, that wouldn't be so terrible. Even now, your phone can download a storage resident copy of Google Maps data for your city, but probably not all the cities nearby in the same storags cache. If you take a road trip, you can update the offline maps cache to keep it current as you travel along.
I'm sure Tesla has some good idea of the various resource limits and possible solutions already, but it may be too early even for them to know exactly what regional data-caching policy they might need.
It seems to me that although the v12 FSD can drive without real-time and precise GPS location, a general region location is probably a very helpful "cheat code" input that gives the model a leg up. Even if it could theoretically come to the right conclusions without any explicit location hint, it would need any more training scenarios.
TLDR: FSD software package customization by region may or may not be needed, but if so that's really not any kind of showstopper, nor any flaw in the approach. It's just another aspect of of deployment logistics if the entire world's good-driving knowledge can't fit into the downloaded model implementation.
Last edited:
JupiterMan
Active Member
My understanding is that the training set is simply Tesla's driving in the situation where the training data is required. This will be human drivers, not FSD for the training set, as such they will have heaps of training data available worldwide and growing hourly. Having less drivers using FSD theoretically helps the training set.
The restriction, as I understand it, is the compute required to train on the data set, not the data itself.
A car in Self driving Tesla in europe will need to drive in multiple countries, i.e. at least 28, and probably more, each with different languages and signage and laws, plus be able to drive on both the left and right hand side of the road.
If they can do this in Europe, then USA should not be any problem at all, I guess the question is not capacity, as training simply adjusts the weights within the existing NN, so is not adding any additional capacity requirements - the question I guess is whether learning roads in say France, will make it more confused when driving in USA can it comfortably learn which data to follow and which to ignore based on location.
My guess is it should be able to do this, in a similar way as it will drive differently on a dirt road to a highway. Tesla do however do different map builds based on location, so its not impossible that different NN weights could also apply based on region.
Elon mentioned in the recent video that they could in theory just push the NN weights to the car as the fleet learns, rather than updating the binary. In a similar way, like you suggest, it may be possible to update the NN based on geography (if needed) also as you drive across borders. Might be a big download though.
Europe has lots of problems to overcome..the sheer abundance of roundabouts ...the tiny narrow roads with cars parked both sides... and drivers coming right at you because they cannot go around a corner without taking both lanes (my main bugbear after decades of driving here)
jeewee3000
Active Member
I've been pondering the V12 across multiple countries question a lot.
If the answer were to be a seperate set of weights for each country/region, then of course you would need a data connection when entering a new geography.
However Elon has hinted that 90% of the driving task is uniform across the globe (don't hit stuff, follow the road, respect signage) and that only 10% is local rules. This would imply the V12 stack should have a worldly driving capacity and is then tuned for separate regions.
Could this be compared to how ChatGPT3 can answer you in any language? I didn't know this at first but when it asks me in English "how can I help you?" I answer something in Dutch and it answers in perfect Dutch. So the NN recognises from the patterns provided (my text input) that it has to use the weights associated with that region.
Tesla could do something similar with FSD v12, therefore incorporating the whole world into the stack at once, but then the model has to be large enough to account for all this, and there must be enough training data. (I guess if the training data contains GPS information that would greatly help the NN to see patterns in "when the GPS points to a location in the UK I have to drive on the left".
So yeah, I guess it's possible to make one big stack to rule all road rules, but the question is if that is the approach Tesla will take. It would take an awful lot of data from all countries.
If the answer were to be a seperate set of weights for each country/region, then of course you would need a data connection when entering a new geography.
However Elon has hinted that 90% of the driving task is uniform across the globe (don't hit stuff, follow the road, respect signage) and that only 10% is local rules. This would imply the V12 stack should have a worldly driving capacity and is then tuned for separate regions.
Could this be compared to how ChatGPT3 can answer you in any language? I didn't know this at first but when it asks me in English "how can I help you?" I answer something in Dutch and it answers in perfect Dutch. So the NN recognises from the patterns provided (my text input) that it has to use the weights associated with that region.
Tesla could do something similar with FSD v12, therefore incorporating the whole world into the stack at once, but then the model has to be large enough to account for all this, and there must be enough training data. (I guess if the training data contains GPS information that would greatly help the NN to see patterns in "when the GPS points to a location in the UK I have to drive on the left".
So yeah, I guess it's possible to make one big stack to rule all road rules, but the question is if that is the approach Tesla will take. It would take an awful lot of data from all countries.
They do care, but that's just the software not working right. NHTSA has stepped in when the software was working right but illegally or dangerously.
Looks like they really separated the two cameras apart, which is interesting.
I wonder how they keep those camera views clean, as the wipers may not get to those areas...
They can have a tube to squirt cleaner, air, hot air to clean the camera. It's an additional cost but probably not too expensive.
A startup discovers Vidar aka structure from motion using just cameras. Everyone uses it, Waymo, Mobileye, Zoox etc. The argument has never been lidar vs cameras, its camera only vs camera+lidar+radar. They all have weaknesses that complement each other. But no Camera is not better than Lidar at measuring distance.
This is there Vimeo channel if you want to see more of their stuff.
NODAR
NODAR is creating the next generation of 3D sensor technology for autonomous vehicles. NODAR’s revolutionary camera-based system provides superior range and…
vimeo.com
DanCar
Active Member
Lidar also advantages at night and during weather events.
diplomat33
Average guy who loves autonomous vehicles
Elon also tweeted a few days ago about how roads are made for vision, so lidar is not needed. He does not seem to understand the purpose of lidar. He seems to be under this false impression that it is lidar or cameras and since cameras can see the world, you don't need anything else. He does not seem to understand the concept of sensor redundancy. The fact is that nobody thinks that the world is designed for lidar only. Of course, roads are designed for vision. He does not understand that lidar is used as a complement to cameras, not to replace cameras.And he is wrong to say that cameras are superior to lidar. Lidar is still superior to cameras in certain areas like range, velocity measurements, low light conditions and direct sunlight. For example, Mobileye's next gen lidar can detect an object just 15 cm high, 170 meters away. It can also measure velocities with an accuracy of 5 cm/s. I doubt camera vision is that accurate.
And lidar works reliably in darkness or in direct sunlight where cameras can fail. Elon is right that the world is made for cameras but he misses that lidar can provide much needed redundancy to camera vision. Lidar is especially reliable for collision avoidance. Having that extra redundancy to avoid collisions is a very good thing when you are aiming for L3/L4/L5. You don't want to rely on vision-only for collision avoidance because if the camera vision fails, you will not be able to avoid the collision. Do you want to deploy L3/L4/L5 with vision-only and have a collision that lidar would have easily prevented? I don't think so. That is why so many of the new luxury consumer cars have a front lidar for AEB and collision avoidance. So it is not about vision being able to see the road, lidar is about adding safety.
They could be significantly reduced by slowing down. Braking distance should never exceed visibility distance, Lidar or no Lidar. And if you're going to add active sensors, choose a wavelength that isn't opaque to fog, not lidar but radar.
Honestly don‘t see your argument. He is technically right in his statement as it’s not Needed. It may add value but that doesn’t mean it’s needed. Can a car run exclusively with cameras? Yes, can a car run exclusively with only LiDAR? No. So by that definition cameras Are superior. All your arguments and rebuttal are fine but that doesn’t make his statement wrong.
MarcG
Active Member
Actually (and I know this because I grew up driving in Europe and taking tests for licenses in two different countries), most signage in Europe other than city names, road names etc. does NOT include words, so Europeans traveling across borders by road do not to learn multiple languages.
Things like speed limits, one way streets, “do not enter”, dead end, etc. are all signs that are common amongst most EU countries and are just based on symbols and numbers, unlike in the US where words are heavily used on such signs (like in the examples above). The one exception that I can remember are stop signs (“arrêt”, “alto”, “pare”, etc.) but even then, I’ve recently seen “stop” being used in many EU countries other than the UK (which is actually not even in the EU anymore).
jeewee3000
Active Member
Can confirm as a EU resident having driven in at least 15 EU countries . Road signage in Europe has huge overlap. If Tesla has a plan for UK (driving on the left) I'm sure they can figure EU out.
The toughest edge case in my opinion is on narrow streets where you have to honk or flash lights before the corner to announce your presence and to listen for honks to know the coast is clear. (Examples to be found in many busy Italian cities, for example).
Similar threads
