But not a traffic light, a boat or a mountain or hill.
Elon Musk | Doer of things on X and sad little man
- Thread starter Pogue Mahone
- Start date
A bigger issue is when cars see something that the system has not seen before. In those cases they can go crazy, and do one of the following things
a) don't detect it at all despite being there, so ignore it.
b) detect it as something different and then do all other estimates wrong (trajectory prediction etc).
There is hope that this can be fixed and there is some research going on, but the problem is extremely difficult, even in academic settings (small clean datasets). Of course, using multi-sensors might help cause something hard to detect in camera, might be easier to detect in LiDAR etc, which is why Tesla's approach to use only camera is stupid. Sure, humans use only 'camera' but until AI perception systems become as good as humans' one, then cars should use other sensors.
Another problem is when you change the city. A car trained in Silicon Valley or Pittsburg will drive very nice there, will struggle quite a bit in New York, would be very bad in Munich, and will be hopeless in New Delhi. While humans tend to generalize far better, and a human trained anywhere will drive relatively well everywhere.
Or the difference between a brown bear and Siberian tiger
That’s it, with Naples’ urban traffic being considered the “grade” in Italy (or the EU, for a common legal framework). I think active learning on the field, more than reinforcement learning plus assisted simulations, will help design “sustainable” algos to minimise accidents… but they will still require a lot of time, then calibaration / fine tuning, and in the end some controlled experiments.
Active Learning will definitely be part of the solution but only a small part of it IMO. It is computationally very expensive to train (multiple cycles of training), expensive to evaluate (you need to do inference in every image in each cycle), and usually does not work well when the datasets get more complex. I kinda developed what is probably the best active learning method for object detection and while it works awesome in VOC (small academic dataset), it works just ok in COCO (medium academic dataset) and essentially doesn’t work much better than random in large industrial datasets.
I think self-supervision is even more important and techniques similar to ChatGPT are promising (Masked Autoencoders for Vision Transformers). Still, the gain there is far less than in NLP, and on LiDAR, I was not able to make these things work when working with large scale data (although I am confident that someone will crack this problem in this or next year). Still a lot of room to improve here and I think it will be the next breakthrough in autonomous driving.
Not sure reinforcement learning will bring much. You can only go so far with simulations and simulations won’t translate well to real world. Now doing reinforcement learning in real world, that would be fun
Active Learning will definitely be part of the solution but only a small part of it IMO. It is computationally very expensive to train (multiple cycles of training), expensive to evaluate (you need to do inference in every image in each cycle), and usually does not work well when the datasets get more complex. I kinda developed what is probably the best active learning method for object detection and while it works awesome in VOC (small academic dataset), it works just ok in COCO (medium academic dataset) and essentially doesn’t work much better than random in large industrial datasets.
I think self-supervision is even more important and techniques similar to ChatGPT are promising (Masked Autoencoders for Vision Transformers). Still, the gain there is far less than in NLP, and on LiDAR, I was not able to make these things work when working with large scale data (although I am confident that someone will crack this problem in this or next year). Still a lot of room to improve here and I think it will be the next breakthrough in autonomous driving.
Not sure reinforcement learning will bring much. You can only go so far with simulations and simulations won’t translate well to real world. Now doing reinforcement learning in real world, that would be fun
Thanks, you are right, of course. My client was looking at (and passed on to me) quadruped robots deployed with model free RL, such as in a paper titled: A Walk in the Park: Learning to Walk in 20 Minutes With Model-Free Reinforcement Learning.
It can tell there's a sizable or moving object, then it guesses what it might be based on predefined data conditions. Meaning if a big sheet of paper or a dozen balloons blow in front of the car, it will slam it's breaks on. There will be edge cases where a car plows down a pram because it thought it was a flag, this kind of thing will happen and people aren't ready for that yet.
For sure there will be edge cases where people get hurt or even die due to automated cars. The recognition software is based on conditions eg size > moving > speed of movement > on the ground/off the ground = most likely object + desired action in current situation.
The more variables you add to those conditions or, the real world adds, the more issues arise.
Yeah definitely, the software is hard to code, which is why I think it's a long way off, it can only adapt to the situations it's been written to adapt to, essentially, what man is capable of thinking about might what happen.
You'd be fecked in Naples. Some Tesla would just come flying through your meal at full speed, pasta and pizza everywhere.
Is this turning into the Tesla thread?
In theory self-driving cars are better for absolutely everyone. But human psychology is finnicky and irrational, and it's hard to know how the companies/government will overcome it.
Even if stats showed definitively that accidents would be reduced, say, 10x if more people switched, the mental exercise of a child being missed by a sensor at a crosswalk would trump it in peoples' heads.
What's crazy is that accidents - and fatal accidents - are both shooting up in the states because the cars are getting so much bigger. There's a great article in the Atlantic (https://www.theatlantic.com/ideas/a...c-vehicles-suv-battery-climate-safety/672576/) highlighting this, including a terrifying study where they sat some guy in the drivers seat of giant truck and he literally couldn't see like a dozen kids that were in front of him, because it's so high up and the hood is taller than them.
Even if stats showed definitively that accidents would be reduced, say, 10x if more people switched, the mental exercise of a child being missed by a sensor at a crosswalk would trump it in peoples' heads.
What's crazy is that accidents - and fatal accidents - are both shooting up in the states because the cars are getting so much bigger. There's a great article in the Atlantic (https://www.theatlantic.com/ideas/a...c-vehicles-suv-battery-climate-safety/672576/) highlighting this, including a terrifying study where they sat some guy in the drivers seat of giant truck and he literally couldn't see like a dozen kids that were in front of him, because it's so high up and the hood is taller than them.
I disagree with this. The object detection/recognition software is essentially just a neural network, in case of images a CNN or Transformer, in case of LiDAR (a point net based, convolutional based on some Transformer). There is no physical model of the world in the system, there is no explicit knowledge on small/large sizes, definitely no speed, and surely no if/else-like decision making. It is as blackbox as it can be.
Then after the objects are detected, they are processed. For example, they can be segmented (another neural network, or maybe the same one), they can be tracked (another neural network), and you can predict their trajectories (probably a neural network but it can also be some classical method like Kalman filters). All blackboxes.
Then based on this information, some decision will be made (probably a neural net). Thing is, when an error is made, it is very hard to know where it happened, cause neural nets are not interpretable.
At the contrary, the code is not the hard part. Neural nets are quite code friendly. But because they do not make much assumptions, you cannot inject 'rationality' in them, so the trend is to make them larger, and train them in more data. But choosing the right data is very hard, and quite expensive. Which is why methods like active learning and self-supervised learning are promising.
I read quite a bit on robo taxis and the concept on the face of it seems good but the reality once you look into it is just terrible. Personal, car-based, and individual transport is not good or efficient. In fact, it's one of the worst methods for transporting people. To then have enough robo taxis to replace the current personal car usage is just unfeasible and would increase congestion.
You're still in the car?
OK, so ask yourself why BMW haven't licened this incredible tech and made an absolute fortune? Perhaps it's because it was autonomous and the reality is the tech still is nowhere near being here.
This isn't the challenge, the challenge is at what point does the vehicle prioritize the safety of the occupant or of the group of children in the road? Perhaps the avoidance route takes you off a cliff? What does the computer choose or do? That's the issue.
People talking of autonomous vehicles forget that a piece of code is written but some over-caffeinated coder at a company who has lobbied responsibility away from their company, will likely make a decision whether to keep you alive or not.
You can be working, reading. Or simply using the time to unwind. Many other things besides driving.
You're still in the car?
The future of urban areas isn't the self-driving car, it's no cars. At least as far as city centres are concerned.
Absolutely, instead of burning VC money chasing something that is basically worthless, spent it on pedestrianizing city centers and creating tram lines, etc.
Disagree. Cars are inefficient cause most of the time they are parked doing nothing. By having cheap robotaxis, people are gonna buy fewer cars, the traffic is gonna get better cause cars won't spend so much time checking for parking etc. So it definitely makes sense, but executing it requires solving autonomous driving, which seems to be harder than initially though.
Besides the point, BMW are pretty shit in this aspect. Waymo (Alphabet) and Cruise (GM) are leading, Tesla is the best of the rest, then there are many other companies ahead of BMW. But most think that this is a two way race between Waymo and Cruise, with Tesla a distant third.
It really is different type of code. Coding neural nets is not that hard, training them is harder. In many ways, the real code is the weights of the network, which are a function of the algorithm and the data, not the Python/C++ script. While there are human-related bugs there, most 'bugs' are gonna be neural net-related so different types of bugs.
A blog post that explains this better from ex Tesla's senior director, who lead Tesla's autopilot until a year ago (though he might have been at OpenAI back then) Andrej Karpathy: https://karpathy.medium.com/software-2-0-a64152b37c35
Everything he said there still stands, just that now the number of weights is in hundreds of billions, instead of millions (some models have reached trillion+ weights). Also, back then (2017) different domains (vision, NLP, speech) used different types of networks, now all are converging to a single type (Transformers).
It depends where. In Europe, I agree. In the US where the cities are far more spread and the roads are bigger, then no, public transport is not the answer. Don't know much about other regions.
Last edited:
Yep. People just don’t want to acknowledge it yet. But electric or self driving cars are no solution, just a slightly smaller problem, at best. The issue is the idea that basically everyone should be able to have his own car and drive it wherever.
Disagree. Cars are inefficient cause most of the time they are parked doing nothing. By having cheap robotaxis, people are gonna buy fewer cars, the traffic is gonna get better cause cars won't spend so much time checking for parking etc. So it definitely makes sense, but executing it requires solving autonomous driving, which seems to be harder than initially though.
Unfortunately, this is incorrect, cars take up space and are inefficient because their occupancy is severely limited. Cars are usually used to transport individuals/two people max per journey. Inefficient use of space. The vast majority of car journeys are "not looking for parking" in the slightest, it is commuting to a location where there is parking i.e. a workplace or office, which has bulldozed a massive area of land, to accommodate stationary vehicles.
I really don't understand how it can be claimed cars "looking for parking" is inefficient but having robo taxis traveling EMPTY to pick up new fares is an "efficient" use of a small vehicle. If these robo taxis are on the road 24/7, being used for journies, like driverless ubers, then they will be more cars on the road because they have replaced the cars that would be parked outside your house (not on the road) and the driverless cars will be on the road constantly.
Besides the point, BMW are pretty shit in this aspect. Waymo (Alphabet) and Cruise (GM) are leading, Tesla is the best of the rest, then there are many other companies ahead of BMW. But most think that this is a two way race between Waymo and Cruise, with Tesla a distant third.
If you think BMW wouldn't move heaven and earth for the chance to get ahead for this trillion dollar business idea because "they are pretty shit in this respect" I dunno what to say.
It really is different type of code. Coding neural nets is not that hard, training them is harder. In many ways, the real code is the weights of the network, which are a function of the algorithm and the data, not the Python/C++ script. While there are human-related bugs there, most 'bugs' are gonna be neural net-related so different types of bugs.
A blog post that explains this better from ex Tesla's senior director (though he might have been at OpenAI back then) Andrej Karpathy: https://karpathy.medium.com/software-2-0-a64152b37c35
Everything he said there still stands, just that now the number of weights is in hundreds of billions, instead of millions (some models have reached trillion+ weights). Also, back then (2017) different domains (vision, NLP, speech) used different types of networks, now all are converging to a single type (Transformers).
You haven't answered my question just shared a blog from an ex-Tesla person. This doesn't answer the question or the moral issue here, that these cars will effectively choose who to kill, which means their code and algorithms are pre-programmed (premeditated) to kill people in a certain event, including the occupants of the vehicle.
It depends where. In Europe, I agree. In the US where the cities are far more spread and the roads are bigger, then no, public transport is not the answer. Don't know much about other regions
If the USA made the decision they could easily start a ten year program to massively increase their public transport networks.
I agree about certain urban centers but if you actually meant restricting ownership of persnonal vehicles I'd say that idea is on the fashy side.
I think there is some misuderstanding here. BMW are not good at self-driving cars (frankly, no German company is), they are not even a serious player.
He was not at Tesla back then, but it does not matter. I see that many people have completely wrong idea on how these programs work, thinking that it is people who write them and make decisions on 'what to do if X happens'. It is very much different to that and that post explains it (the post itself does not even talk about self-driving cars).
To answer the question, it likely depends on the data. Considering that most people try to save themselves fast, the data is gonna be biased towards that. Of course, you can hack it by oversampling the data where the driver decides to save the kids and kill themselves (or weight those samples more), but I do not expect companies doing that. What will happen is oversample cases where both survive, which should be the goal
I disagree. Cities are build very different there, and are far sparser for an efficient public transport.
If aerospace regulation is to be taken as precedent I think it would be very hard to get this adopted when the ultimate decision entity is essentially a black box and not a clearly defined state machine.
Agree, the legal framework is gonna be a clusterfeck.
I also think that ultimately it will get adopted. Never in our history, we have declined to use technology.
BTW, aereospace industry is exploring the usage of modern AI too. I actually interviews and had an offer from Airbus last month, where one of the things I would have worked on would have been autonomous driving for airplanes. They seem more interested in interpretability than any self-driving car I am aware of, though.
Legislation issues.
The sorts of legislation issues that mean that cameras for wing mirrors arent yet a widespread thing because only a very small handful of countries have legislated for them. Cameras for wing mirrors improve how slippery a car is through the air which is a major assist for improving EV battery range. Have a look into that, its a whole thing.
The tech isnt there yet but its far far further along than people think. Its not just the cars that need to be ready for level 4 or level 5 Auto driving, roads, signage etc all need to be upgraded. Upgraded in terms of things that will talk to the cars.
Car design is roughly 3-4 years ahead of what makes it to market.
There are currently new trials of autonomous transport busses and cars taking place in various cities. There have been trials that have finished and new ones start.
My example was something from 4 years ago.
Ask yourself if you think BMW have just stopped there.
Last edited:
Damn straight!
I was promised hover boards and jetpacks and flying cars by now. All I've got is a PS5 and a fecking Smart phone.
I'd much prefer a hover board.
Can we get back to Musk being a cnut?
At least that's what most would agree on.
At least that's what most would agree on.
I am not however, "wasting time behind the wheel." I honestly don't know what point you think you are making.
He's definitely a something and, whatever that is, he's a complete and utter one of it.
Trying to figure out who he reminds me of has been annoying me for ages, and I've finally got it.
It's jedward, he reminds me of jedward.
He's given his mother a lovely pearl necklace.
Trying to figure out who he reminds me of has been annoying me for ages, and I've finally got it.
It's jedward, he reminds me of jedward.
Last edited by a moderator:
No worries
Trying to figure out who he reminds me of has been annoying me for ages, and I've finally got it.
It's jedward, he reminds me of jedward.
Are we the baddies?
1. No they won’t
2. 99% of drivers are sober already. Alcohol is the problem. Not the car.
3. Trains and buses exist
4. This ‘leasing’ system already exists
5. Parking issues will scale. The fewer spaces we need, the more that people can reclaim streets.
6. The traffic jams point may have a small point of value, but weight of traffic means it will Never be solved by a vehicle.
7. Driverless cars crash
8. If you think driverless car companies will be on the hook for killing people after you elect to let it drive itself… you’re mad. It will still be your fault.
None of what you believe is really tied to reality. I used to believe most of it. It falls over at the first point of sensible interrogation though. I’m almost all directions.
None of it works without foolproof tech and 100% adoption. We will never have either.
It's complete insanity. There's no feckin way I'll ever trust it.
But what has he done?
I'm no expert in the subject, but I'd think self driven cars are not a long term solution for the transportation issue. IMO big cities could and should have some self driven taxi system that drastically reduce issues with traffic, accidents, parking space requirements, etc. and for uptown/small towns plus long distances a strong transportation network with priority/exclusive roads (ideally trains) should do the trick. That leaves particular cars (self driven or not) mainly for the last mile altough they can still be usable for long distances at increasing costs, since they'd have to cover most of the share for infrastructure, accident risk costs, etc. Well, either that or the Jetsons/Futurama solution
Trying to figure out who he reminds me of has been annoying me for ages, and I've finally got it.
It's jedward, he reminds me of jedward.
Ah now Jedward are alright these days