-
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.
These sorts of situations are probably pretty easy. You assume anything that could be bigger than a squirrel could be a human. First rules are don't hit stuff and stay out of the way of moving stuff that could hit the car. The car doesn't need a high probably ID on what the object is to follow the most important rules. Avoiding contact with stuff in the environment avoids almost all of the worst situations.
More challenging is what to do when the ghost or fire hydrant remains in the road blocking the path. I assume that Waymo level 5 is set up for a remote human to join the car and decide on an action.
The problem is that if you assume that everything is high-risk and could do anything within the bounds of physics then the car sits paralyzed. So then you dial it back a bit and you get lots of awkward, surprising phantom braking (just like current Autopilot). Dial it back more and you get a smooth ride with no surprises but then 0.1% of the time you get a false negative and you hit something -- like Autopilot sailing into the back of a fire truck.
People don't realize how many risks they take, as human drivers, every moment of their drive. There could be a pedestrian between those two parked cars ready to come out into the roadway. That pedestrian walking on the sidewalk probably isn't going to step into the road suddenly, but maybe he'll accidentally trip and stumble into the road. Any one of those car doors of the parked cars could open suddenly. A self-driving system built to always protect against the worst case will never be able to do anything. This is why self-driving is so hard -- it's not that it's so hard to not hit anything ever -- especially if you have lidar plus cameras plus radar this is a tractable problem. But it's hard to not hit anything while still driving in a reasonable way, taking the kind of reasonable risks that humans take every minute of every day.
And occasionally, it means taking a normal, reasonable risk and having it turn out badly. Just like human drivers. Somehow we forgive humans when this happens, but will we be so forgiving of the robots and the companies that built them?
I'd say yes if it's plausible that a human couldn't have avoided it. But I think this will be better than humans in the long run. The processing can concentrate on everything at the time, whereas a human only have 2 eyes and ears.
1-2 months is old? This is what they are doing right now.
I agree it will be better than humans, very soon even. But there is an emotional, irrational aspect to the way people react when machines screw up -- even if the machines save lives overall, how would you feel if it were your loved one that died as a result of software error? Maybe even an error that a human would not have made -- software will make different mistakes than humans make; fewer mistakes overall, but they may likely be different mistakes, especially at first. That's the tough part.
Also with respect to humans "only" having 2 eyes and ears, don't forget that humans also carry around a very energy-efficient supercomputer that still far exceeds anything humans can fit into the trunk of a car in raw processing power, plus it has had millions of years to work out the kinks in its software. Self-driving systems must compensate for lack of this kind of cognition with far superior sensory systems. (This is why lidar is so important.)
1-2 months is old? This is what they are doing right now.
In the age of the internet where news is posted within 5 minutes, it's "old". That video was first posted and discussed on this very forum a couple months ago.
1-2 months is old? This is what they are doing right now.
I expect that Karpathy has had his public speaking privileges revoked.
You just need two or more sensing systems that are highly uncorrelated for ambiguous(lower probability) IDs. Tesla seems to be to do this with radar. The problem may be that radar isn't enough of a generalist to compliment vision.
There's an object in the road that it ID as a large branch ( a generic important obstruction) but at a low probability. Lidar sees no obstruction at a high probability. The car drives on. It's a shadow.
AFAIK you don't train two different systems to recognize stuff and then compare results. One system is trained with multiple input types.
The argument for camera as the only generalist may be that there is sufficient spectrum sensed to effectively provide multiple ways to sense objects. I presume NN output needs to be explicitly converted to probabilities to make driving decisions.
A plastic bag floating across the freeway is quite a bit larger than a squirrel. Stop for that when going 65+mph and you risk killing people. The example I used was just that: an example. There are a great many edge cases where a human can generally do a good job of identifying the object as something somewhere between safe to hit to something that must be avoided at all costs that are much harder than anything in ImageNet. Usually involving "seeing" an object via context with other objects around it, sometimes even involving the way it moves rather than based on a single image.
strangecosmos
Non-Member
This podcast was a fascinating listen:
#45 - Building Dota Bots That Beat Pros - OpenAI's Greg Brockman, Szymon Sidor, and Sam Altman - Y Combinator
The big difference between the topic of the podcast and the topic at hand is that playing Dota is a pure software problem whereas self-driving is a real world robotics problem. The number of Dota games you can run is limited only by your available computational power, whereas real world test driving is limited by economic, physical, and safety factors. Sure, we can do simulations, but simulation is not limited by computational power — it’s limited by real world driving data.
What’s exciting about what Tesla is doing with Hardware 2 in 170,000+ production cars and massive data labelling is that it’s training its systems on an unprecedented volume of data. There is no way to estimate Tesla’s rate of progress because nothing at this scale has ever been tried before.
One general property of deep learning systems is the potential for explosively fast improvement (providing the right conditions are met). In the podcast, the OpenAI team describes how their Dota bot went through a major improvement literally overnight.
Could this sort of ridiculously fast improvement happen with self-driving cars? Such that they could go from a janky demo today to better than the average human in 18 months.
This possibility is one reason why I have a hard time discounting Elon’s prediction that Level 5 autonomous driving will be superhuman by the end of 2019, even though the state of the technology — not just Tesla’s, but Waymo’s, Cruise’s, and everybody’s — seems so far away from that today.
It’s also possible that superhuman Level 5 autonomy will take more than 30 years, or it could turn out that driving requires human-level general intelligence and therefore self-driving cars are impossible (unless one day we decide to enslave sentient computers).
The more exciting scenario is the one where self-driving cars take an OpenAI bot-like exponential trajectory of improvement and move from barely working prototype to commercially viable product so fast that the process feels almost like an instantaneous flash. Elon has been arguing that this is what will occur. I don’t know that he is right, but I also see a lot of people dismissing this possibility out of hand because of Elon’s past tardiness or because they simply assume progress will occur linearly — rather than arguing against the possibility for principled AI/robotics reasons. The tardiness counterargument is fair, but even if you double Elon’s timeline, 3 years is still pretty damn fast. Progress might happen linearly, but it also might happen exponentially, and I would prefer us to talk about the reasons why rather than simply assume one or the other.
For me, the 2018.10.4 update was a huge confidence boost for the exponential progress scenario. There was a seemingly instantaneous improvement (from one update to the next) in certain modular driving tasks, with one owner claiming that after the update their car could do lane changes better than they could. Autopilot was also able to recognize driveable road much better than before.
The big question is whether we’ll see that kind of update cadence continue or even accelerate. If in 5 months we go from 2018.10.4 to Level 2 autonomous on the highway with no driver input plus some other as-yet-unknown advanced features, that will boost my confidence that the end of 2019 timeline (or roughly around then) for full Level 5 is achievable.
#45 - Building Dota Bots That Beat Pros - OpenAI's Greg Brockman, Szymon Sidor, and Sam Altman - Y Combinator
The big difference between the topic of the podcast and the topic at hand is that playing Dota is a pure software problem whereas self-driving is a real world robotics problem. The number of Dota games you can run is limited only by your available computational power, whereas real world test driving is limited by economic, physical, and safety factors. Sure, we can do simulations, but simulation is not limited by computational power — it’s limited by real world driving data.
What’s exciting about what Tesla is doing with Hardware 2 in 170,000+ production cars and massive data labelling is that it’s training its systems on an unprecedented volume of data. There is no way to estimate Tesla’s rate of progress because nothing at this scale has ever been tried before.
One general property of deep learning systems is the potential for explosively fast improvement (providing the right conditions are met). In the podcast, the OpenAI team describes how their Dota bot went through a major improvement literally overnight.
Could this sort of ridiculously fast improvement happen with self-driving cars? Such that they could go from a janky demo today to better than the average human in 18 months.
This possibility is one reason why I have a hard time discounting Elon’s prediction that Level 5 autonomous driving will be superhuman by the end of 2019, even though the state of the technology — not just Tesla’s, but Waymo’s, Cruise’s, and everybody’s — seems so far away from that today.
It’s also possible that superhuman Level 5 autonomy will take more than 30 years, or it could turn out that driving requires human-level general intelligence and therefore self-driving cars are impossible (unless one day we decide to enslave sentient computers).
The more exciting scenario is the one where self-driving cars take an OpenAI bot-like exponential trajectory of improvement and move from barely working prototype to commercially viable product so fast that the process feels almost like an instantaneous flash. Elon has been arguing that this is what will occur. I don’t know that he is right, but I also see a lot of people dismissing this possibility out of hand because of Elon’s past tardiness or because they simply assume progress will occur linearly — rather than arguing against the possibility for principled AI/robotics reasons. The tardiness counterargument is fair, but even if you double Elon’s timeline, 3 years is still pretty damn fast. Progress might happen linearly, but it also might happen exponentially, and I would prefer us to talk about the reasons why rather than simply assume one or the other.
For me, the 2018.10.4 update was a huge confidence boost for the exponential progress scenario. There was a seemingly instantaneous improvement (from one update to the next) in certain modular driving tasks, with one owner claiming that after the update their car could do lane changes better than they could. Autopilot was also able to recognize driveable road much better than before.
The big question is whether we’ll see that kind of update cadence continue or even accelerate. If in 5 months we go from 2018.10.4 to Level 2 autonomous on the highway with no driver input plus some other as-yet-unknown advanced features, that will boost my confidence that the end of 2019 timeline (or roughly around then) for full Level 5 is achievable.
Last edited:
How do they train their systems? Do they upload video to people who label it? Or is it somehow processed locally? I thought you needed to label stuff, so someone should get an (preferably uncompressed) video or a series of images.
caligula666
Member
Yeah, that’s a good question. Labeling indeed needs to be very massive. Is it the reality though? With 170k cars out there, how many FTE are needed to label things properly? Or is it leveraging other data such as maps? But then not really accurate?
Then, if Tesla was to truly leverage all of its recorded data (unclear how much is actually recorded btw), how much computing power would that require in their DC or at a cloud provider? Anybody has any guess there?
Lastly, you say that Waymo, much like others, is very far away from L5, but you link an article that doesn’t quite say that. Agreed it is very L4’ish, but what makes you pack Waymo with Tesla here? Tesla seems at this point very far from a L4. As for L3, we will see in August what gets released, but my bet is that this will remain a L2 for a while (I.e. until at least 2020), as I can’t see the logic of what would need a series of huge steps improvements when anything we have seen to date was slow and linear. I’m sure more stability in the AI team will greatly help though.
strangecosmos
Non-Member
Images are uploaded to Tesla’s servers and then manually labelled by Tesla employees. Check out Andrej Karpathy’s talk.
strangecosmos
Non-Member
The article talks about how Waymo’s cars struggle with left turns. We actually have no idea what the state of Tesla’s full self-driving development is because Tesla is so secretive about it and has hardly released a shred of information since its demo video from October 2016. Comparing Waymo’s full self-driving prototypes to Tesla’s Autopilot is comparing apples to orange juice.
When Waymo first started using deep learning, its pedestrian detection improved 100x in a few months. Cruise’s disengagement rate recently improved 14x in a year. Since collecting and labelling real world driving data seems like the most important factor in the rate of progress, and since Tesla has access to the most real world driving data, my hypothesis is that Tesla is making progress at a faster rate than any other company. This remains to be seen, however.
Have you read this? Karpathy argues, that developing AI for games is much easier than AI for real world.
AlphaGo, in context – Andrej Karpathy – Medium
Don't you mean L3 on the highway with no driver input? Autopilot is already L2.
Similar threads
