Thanks! I posted that video earlier, but somehow I missed the part how they actually collected the images.
-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
Firmware 9 in August will start rolling out full self-driving features!!!
- Thread starter diplomat33
- Start date
-
- Tags
- 9.0 Tesla Inc.
Thanks! I posted that video earlier, but somehow I missed the part how they actually collected the images.
strangecosmos
Non-Member
The SAE levels of automation don’t capture a lot of important distinctions. Model 3s can autonomously drive into in your garage. Technically, that makes them Level 4 in your driveway. The definition of Level 4 doesn’t include any stipulations about which driving modes don’t have a human in the loop.
Similarly, Audi’s planned Traffic Jam Pilot system is Level 3, even though it’s much less capable than Autopilot today.
Adaptive cruise control and lane keeping on the highway is sufficient to be classified as Level 2, but insufficient to drive on the highway without driver input. Driving on the highway, on ramp to off ramp, with no driver input required a much more extensive set of features. However, both are still Level 2 systems.
I think Enhanced Autopilot will most likely remain Level 2.
strangecosmos
Non-Member
Yeah, that’s a great article. There is a big difference between pure software applications of AI and real world robotics applications. The key similarity between Go and Dota is access to a perfect simulator — the game itself. The evaluation is also clear, and trial-and-error is harmless since it doesn’t occur in the physical world.
However, Dota, unlike Go, is not fully observed. The action space is not discrete. I don’t think the game is fully deterministic either. Games are much longer than Go too, comprising a lot more actions.
Like Go, there is a lot of data on human play with Dota. However, Tesla also has access to a lot of data on human driving. So that is one similarity between Go, Dota, and driving.
The main difference with Dota and driving is that you can train an AI on an essentially unlimited amount of Dota playing with enough time and compute power. With driving, the limitation is real world driving. What’s exciting though is the scale that Tesla is operating at. Tesla’s Hardware 2 fleet is driving as many real world miles per day as Waymo is doing in simulation. Soon it will drive more, as the fleet continues to grow.
I would argue that feature is L2 or at best L3, because you are required to supervise the vehicle while using Summon. You're just allowed to supervise from outside the car. Even in this limited scope (driveway and garage, low speed), there are a number of situations that the car cannot handle at all -- objects lower or higher than the ultrasonics, posts, things that don't reflect ultrasonics well -- such that you would need to create a much narrower scope consisting of objects only at the proper height and made of the proper material to begin to call it L4.
And this is why Smart Summon as initially described is probably never going to happen - it implies unsupervised autonomy. And I don't think the current hardware is ever going to allow unsupervised autonomy in any form; you're always going to remain responsible for the safety of the vehicle and will be required to remain attentive at all time. This means L2. They might possibly get some geofenced highway L3.
People with the FSD option will in fact get something that drives itself on local roads, stops at stop signs, etc. But they will have to supervise it the entire time, making it still an L2 system. I don't think the "park seek" mode will ever materialize even for people who paid for FSD, because that's unsupervised.
Edit: So basically, my current thesis, supported by recent changes to the produce descriptions on the website, is that EAP will be (very good) L2 on the highway, and FSD will add (pretty poor) L2 on (some) local roads, and there may be geofenced, weather-limited L3 on the highway for FSD purchasers as well. If they can pull off true L3 on the highway with the current hardware that will be a triumph. (Except it will still fall far short of the promises.)
There has to be image
Only images, no video sequences? To capture for example a blinking turn signal, it has to actually recognizing the motion /sequence of blinking. Or else it could be brakes with one of the lights dead. The sequence from red to yellow to green light etc. Or the reverse. If it sees from red to yellow when you drive into one road, it might abruptly stop believing it goes from yellow to red, and not yellow to green.
Are these hand programmed and sprinkled with trained still images?
Only images, no video sequences? To capture for example a blinking turn signal, it has to actually recognizing the motion /sequence of blinking. Or else it could be brakes with one of the lights dead. The sequence from red to yellow to green light etc. Or the reverse. If it sees from red to yellow when you drive into one road, it might abruptly stop believing it goes from yellow to red, and not yellow to green.
Are these hand programmed and sprinkled with trained still images?
I agree with you that the SAE levels don't capture a lot of specific distinctions which is why different people have slightly different interpretations of what the levels actually mean in the real world. I would disagree with you that summon is L4 in your driveway. With Summon, the car is not self-driving at all, you are directly moving the car by remote control using your phone. It's no different than when you drive your a little toy car with a remote control. That's not L4.
I would also point out that the SAE levels don't say anything how to monitor driver engagement. Just because a system tracks the driver's eyes to gauge engagement does not automatically make it L3. For example, I would argue that GM's Supercruise, while hands free, is still L2.
I would disagree that it would still be L2. L2 is defined as a system where both speed and steering are automated by the car but the driver does everything else including monitoring the environment. L3 is defined as a system where both speed and steering are automated AND the car also monitors the environment but the driver may need to take over (hence still needs to supervise). So L3 can still require driver supervision. Just because the driver may need to supervise does not automatically bump it back down to L2. In your scenario, if the car is driving itself on local roads and able to stop at stop signs, traffic lights etc, then it is monitoring the environment hence by definition it is L3, not L2.
For reference again:
Well, the definition does say that L2 systems "use information about the driving environment". And if the car is doing steering and accel/decel what is left to automate? But I think what we can all agree on is that the SAE level definitions are abiguous and we could argue all day where any given real-world system should land between L2 and L3 in particular. I'm not really interested in that argument. I'm interested in whether you will ever be allowed to read a book while letting the Tesla handle the driving, or truly "smart summon" it out of your garage and then let it park itself unsupervised at your destination. Or, of course, the holy grail being to let the Tesla be an autonomous taxi in a ride sharing network (which implies you don't need a driver in the car at all), or summon itself across the country to come and get you.
Define L2/L3/L4/L5 however you like. I'm saying that AP2/AP2.5 vehicles are going to have very limited unsupervised capabilities, if any at all, despite the promises made about this.
Joe F
Disruption is hard.
Quite possibly true, however, is anyone seriously considering using their Tesla that way? I for one certainly would not, and doubt anyone who values their vehicle would either, even if ever made available, which I too doubt.
Red herring. Tesla promised it, they need to deliver regardless of what you think is useful or prudent.
As a software engineer and lead developer on an mobile app that is about to deploy in 7 days that services over 2 million individuals that uses multiple neural net models to scan medications.
It is 100% Jargon. Look up what jargon is. Ask any machine learning engineer or data scientist what software 2.0 is and they will look at you wrong. Now ask them what data gathering, data labeling, building models and you will have a hour long conversation.
In-fact you find no mention of "software 2.0" in google up until November 2017 when Andrej wrote his blog.
That is the definition of jargon.
data gathering and data labeling is machine learning 101. Software have been compromised of just NN models for a long time. People who think its something new and cutting edge are simply misinformed.
I'm not surprised remember "fleet learning" and "shadow mode"?
Last edited:
Blader, you did not respond to the message you quoted. You responded about the term "Software 2.0". Why don't you respond about whether the processes are the same or different for machine learning vs writing code? Obviously ML also involves some hand-written code, but it's a smaller chunk and you can have many "releases" of an ML system without touching a single line of hand-written code, so the business and engineering processes involved are clearly different. Similarites, yes; overlap, yes; but also there are notable differences.
The SAE absolutely does capture all the distinctions. The actual SAE document is like 20 pages long or something.
First of all the feature set for Level 4 is NOT on-ramp/off-ramp, or garage.
The Model 3 does not autonomously drive into your garage because you are controlling and monitoring it.
https://pbs.twimg.com/media/DhbmR7GXkAMd6IP?format=jpg
The SAE clearly defines the difference between the levels and i don't understand why people keep messing it up.
It has nothing to do with on-ramp/off-ramp.
The feature set for Level 3 is the autonomous system monitoring the environment and responding appropriately and alerting the driver ahead of time if they need to take over. The driver would not need to monitor the environment or the car anymore and can then read a book, watch TV/MOVIE, text, etc. But they can't fall asleep because they need to be ready to regain control in sufficient time (6-10 seconds) in-case they are called upon for fallback. That is what defines a level 3, nothing else matters.
The feature set for Level 4 is the autonomous system monitoring the environment and does not require the human driver to retain fallback control of the driving task in any conditions, even in an emergency. It will handle the emergency and come to a safe stop. That is what defines a Level 4, not on-ramp/off-ramp.
on-ramp/off-ramp is just another operation domain. just like parking lot, traffic jam, dense traffic, full highway speed, city, weather, suburb, etc.
In conclusion:
Level 2 - Hands off (ex: Supercruise)
Level 3 - Eyes off (ex: reading book, watching movie aka Audi Traffic Jam Pilot)
Level 4 - Mind off (ex: sleeping aka Waymo)
But the SAE makes it very clear in its documents.
Last edited:
Yes, agreed.
Personally, I think being able to read a book while the car cruises on the highway is definitely doable with AP2. I also think having the car get out of a parking spot and pick you up at the entrance of a store unsupervised, is probably doable with AP2. Now, the ride-sharing network where the car drives around town with nobody in the driver's seat, no, I honestly don't think that will happen on AP2 any time soon.
Joe F
Disruption is hard.
True, but I'm stating my use case. Many features are released by many CO's which have no use in the real world. See Microsoft: Clippy.
I'd rather see resources put towards fixing real issues, and releasing features which will actually be used. Personally, I'm not a fan of the sharing economy system, especially as applied to a personal vehicle.
I responded to the retort to my statement that "Software 2.0 is complete jargon" is not jargon and i proved without a shadow of doubt that it is. Its just a hype term recently used (starting from Nov 2017 according to google) to describe processes already known and standard in the industry to make it seem to be something new. Which is per for course for Tesla.
I'm not arguing that the processes are the same. But that the process are 101 and basics. There's no need inventing a hype umbrella term. Everyone associated with Tesla marveled at Andrej presentation, but any data scientist or machine learning engineer would shake their head at how basic that presentation is.
Joe F
Disruption is hard.
I would think the front radar could detect road debris. And Tesla could probably design the camera vision to recognize road debris too although I am sure that would be a challenge.
Radar's going to have trouble with some of the smaller debris, and certain materials. Cameras should, in theory, do as well as humans. But detecting and correctly categorizing the threat level of debris is only the first part of the challenge, arguably the easy part. Reacting quickly and correctly is the hard problem here. It is very easy to react to road debris in a way that causes rather than prevents accidents.
I agree with you that the SAE levels don't capture a lot of specific distinctions which is why different people have slightly different interpretations of what the levels actually mean in the real world. I would disagree with you that summon is L4 in your driveway. With Summon, the car is not self-driving at all, you are directly moving the car by remote control using your phone. It's no different than when you drive your a little toy car with a remote control. That's not L4.
I would also point out that the SAE levels don't say anything how to monitor driver engagement. Just because a system tracks the driver's eyes to gauge engagement does not automatically make it L3. For example, I would argue that GM's Supercruise, while hands free, is still L2.
I would disagree that it would still be L2. L2 is defined as a system where both speed and steering are automated by the car but the driver does everything else including monitoring the environment. L3 is defined as a system where both speed and steering are automated AND the car also monitors the environment but the driver may need to take over (hence still needs to supervise). So L3 can still require driver supervision. Just because the driver may need to supervise does not automatically bump it back down to L2. In your scenario, if the car is driving itself on local roads and able to stop at stop signs, traffic lights etc, then it is monitoring the environment hence by definition it is L3, not L2.
For reference again:
People keep saying the SAE is ambiguous, but that's not true. The problem is that they keep using just that one screenshot chart.
Here is the actual full SAE document
https://wiki.unece.org/download/attachments/40009763/(ITS_AD-10-08) SAE_J3016_Taxonomy and Definitions for Terms Related to Driving Automation Systems.pdf?api=v2
Its 30 pages long. Its very comprehensive.
No L3 does NOT require driver supervision. It requires driver as safety fallback after appropriate timed request (up to 10 seconds according to the industry). That's a big difference. Again the 30 page document goes into detail about this.
Thanks for the link. If you are expected to be the safety fallback and take over within 10 seconds if necessary, then you still have to be paying attention. So maybe you are not directly supervising the system, but you are still paying close attention to what it is doing. How could you be the safety fallback if you are not paying close attention? That's what I mean by "supervise". The driver is still expected to watch what the car is doing so as to intervene if needed.
Similar threads
