Welcome to Tesla Motors Club
Discuss Tesla's Model S, Model 3, Model X, Model Y, Cybertruck, Roadster and More.
Register

Karpathy talk today at CVPR 2021

This site may earn commission on affiliate links.
I looked through for the source because from my knowledge of lidar sensors, presenting a FOV like that can be misleading on the resolution and coverage of the sensor. I found you show a birds eye view of the older sensor, but didn't for the new one. At 1:30:37 it shows a similar view and you can see the blind spots of the sensor.
Waymo

View attachment 676795

What blind spot? You are kidding right?
 
There’s a lot of theoretical perspective in this thread and not enough practical perspective IMO. Specifically on the topic of sensor fusion and what is “best”. In the theoretical view, sensor fusion should always be better. The more types of sensors, the better. The more number of sensors, the better. The more resolution/bandwidth the sensor, the better. In theory. But, Tesla is working to solve the problem ASAP, which means they have limited resources and choices to balance. AK even admitted that they “could” go dig into why their radar stack is giving them noisy/unreliable data, but that their (limited) engineering resources are better spent on the richer sensor stack. They have to make decisions that balance the very finite resources (both hardware, compute, and labor) being applied in the best way possible to get to FSD that works better than humans by xxx% ASAP. So while 24x 4K cameras at 120Hz fused with hi-res & hi-bandwidth LiDAR fused with hi-res RADAR fused with ultrasonics theoretically gives awesome data, we don’t practically have the HW to process that in our cars today, nor the backend to train with that amount of data. Trying to solve FSD with that HW stack would take a long, long, long, time. Today we have 8x 1K cameras at 36Hz and Tesla thinks that’s enough to get there. Once they “get there”, maybe at 10x human driver safety, the NN continues the march of “9s” and gets to 15x or 20x. Then Tesla might update the sensor suite (higher resolution, maybe adding back in xxDAR) and take it to 100x or 200x. Maybe basic FSD offers 15x and upgraded FSD (more money for better sensors) offers 100x.
 
Practically speaking no one has proven the safety of any autonomous vehicle to be greater than the average human let alone 10x safer. For the cheapest solution to end up being the first successful implementation would be very surprising and maybe unprecedented (are there any examples?).
Tesla's use of 18650's? It was the cheapest solution at the time to get a long range battery and that's what Tesla went with and got there first. The others focused on "better" prismatic cells, which limited the size of the battery pack they can add to the vehicle and took longer to get there.
 
Practically there are also constraints on AI training time. If it takes 2 weeks to train a vision neural network and 2 weeks to train a radar neural network, you can drop radar and instead train 2 vision neural networks in 2 weeks with slightly different parameters and use the better one. Or ship both vision networks and a/b test or average the results (aka a model ensemble).
 
To convey why everyone should be fired up about SSL, I share the words of AI luminary Yann LeCun:

"How is it that many people learn to drive a car fairly safely in 20 hours of practice, while current imitation learning algorithms take hundreds of thousands of hours, and reinforcement learning algorithms take millions of hours? Clearly we’re missing something big.

It appears that humans learn efficiently because we build a model of the world in our head. Human infants can hardly interact with the world, but over the first few months of life they absorb a huge amount of background knowledge by observation. A large part of the brain apparently is devoted to understanding the structure of the world and predicting things we can’t directly observe because they’re in the future or otherwise hidden.
As you say human infants learn by observation. It does take us 16 years to get to the point we are deemed ready to drive. In that time we have mastered walking, distance perception, cognition and any number of other things. We may have learned much about driving by observation in that time too. So it's around 140,160 hours of learning to be human enough to be the average human driver. If you count sleep cycles which are probably important too.
 
What blind spot? You are kidding right?
The obvious circle that surrounds the vehicle that is not from the occlusions by other objects, but rather simply the sensor not covering that area (the 360 sensors have limited vertical FOV so it typically is not used to cover the near field even if it was possible to do so). This circle is present for basically any of these roof mounted sensor solutions.

Some use ultrasonics to cover that blind spot, but it appears Waymo uses other perimeter lidar to cover it.
https://www.eetimes.eu/abandoned-technology-maps-lidar-blind-spot/
Waymo%2Bcar.JPG


In gen 5, it seems they added cameras to cover all those spots:
iPace-lineart-sensor_calloutv2_03022020-01.png


So basically the opposite of Tesla, which relied a lot on vision from the start, but Waymo is now adding more and more vision to their suite.
 
Last edited:
  • Informative
Reactions: scottf200
Tesla's use of 18650's? It was the cheapest solution at the time to get a long range battery and that's what Tesla went with and got there first. The others focused on "better" prismatic cells, which limited the size of the battery pack they can add to the vehicle and took longer to get there.
When the Roadster came out it was the most expensive EV ever sold.
My point is that usually the first versions of products are very expensive. Also, there are usually functional prototypes built first.

The obvious circle that surrounds the vehicle that is not from the occlusions by other objects, but rather simply the sensor not covering that area. This circle is present for basically any of these roof mounted sensor solutions.
That's why they put lidar on the corners. I think the Zoox and Cruise purpose built AVs also have sensors on all four corners.
 
Last edited:
This slide gives us an example of the scale of data Tesla can collect for SSL in 4 months. 6 billion objects! 1 million video clips! That's an average of about 1 clip from each post-HW1 car during that 4 months.

To people who say Tesla has access to more data than they can usefully exploit... I say pah! Their average in this instance was ~1 video clip per car.

This corroborates the finding that pretty much anyone who monitors their Tesla notices that it uploads a lot of megabytes. The scale of the fleet is critical to making SSL projects like this work.

View attachment 676817
This does somewhat contradict the quote you posted earlier. According to that quote, SSL should result in the need for fewer clips, not more.

On the other hand, we don’t yet know how many clips are “enough,” so it may only take 1 million clips instead of 10 million.
 
When the Roadster came out it was the most expensive EV ever sold.
My point is that usually the first versions of products are usually very expensive. Also, there are usually functional prototypes built first.
The Model S then came out 4 years later in 2012 at half the price and still with 200+ miles of range. The others didn't get to that until 4 years after the Model S with the Chevy Bolt in late 2016 (which sold at compliance car volumes, volumes which Model 3 immediately smashed when it came out shortly after). The rest didn't come until a whopping 7 years later in 2019 (with cars like the iPace, e-Tron, 60kWh Leaf, Kona). But the focus on that comment is more on the tech: they used the cheapest/simplest battery tech available and got to a workable solution first, rather than using more advanced/expensive battery tech like others that came out years later.

And to extend this, they took what they learned in working with commodity cylindrical cells to extend to 2170s and then now to 4680s. We may see the same thing here, that they take what they learn with using relatively cheap camera sensors and extend to more advanced ones. But first the immediate goal is to get to a workable solution first.
 
Last edited:
Practically speaking no one has proven the safety of any autonomous vehicle to be greater than the average human let alone 10x safer. For the cheapest solution to end up being the first successful implementation would be very surprising and maybe unprecedented (are there any examples?).
Well, without data to show FSD is better than humans, I think getting regulatory approval will be arduous. I believe showing FSD in xx% better than humans makes this increasingly easier, because at some point it’s a no brainer…My opinion is that L5 FSD, when approved, will be at least 10x safer than human drivers.
 
The Model S then came out 4 years later in 2012 at half the price and still with 200+ miles of range. The others didn't get to that until 4 years after the Model S with the Chevy Bolt in late 2016 (which sold at compliance car volumes, volumes which Model 3 immediately smashed when it came out shortly after). The rest didn't come until a whopping 7 years later in 2019 (with cars like the iPace, e-Tron, 60kWh Leaf, Kona). But the focus on that comment is more on the tech: they used the cheapest/simplest battery tech available and got to a workable solution first, rather than using more advanced/expensive battery tech like others that came out years later.

And to extend this, they took what they learned in working with commodity cylindrical cells to extend to 2170s and then now to 4860s. We may see the same thing here, that they take what they learn with using relatively cheap camera sensors and extend to more advanced ones. But first the immediate goal is to get to a workable solution first.
The Model S was second only to the Roadster in cost! Are the 18650 packs actually cheaper? They are more energy dense and higher performance for sure but the BMS is complicated and the assembly costs are high which is why Tesla is now going to larger cells (after figuring out how to match the performance of smaller cells). Tesla started with a very expensive EV and worked towards making cheaper EVs. What they're doing with FSD isn't analogous at all.
This is all besides the point since EVs have existed for about 150 years. I'm talking about inventing something entirely new like a computer or cellphone or any number of automotive features that are initially very expensive and later become standard. Right now it's not clear if even the most advanced AV prototypes can match the performance of an average human.
 
I don't think it is that simple. There is a computing cost to integrating disparate sensor data. That is the trade-off I was thinking of.
yes, it will cost processing time, but I'm not one to limit the cpu power of systems that need to get real work done. (at work, they love to cost reduce even the cpus; sometimes we can bump up the core count but then, the beancounters force us down with some lower memory amount. its insane who much fighting there is just to get the board power you think the project will need. sigh..)

so, the amount of processing is irrelevant. you size the systems to be as good as you need, with some headroom to spare. and you spec your test/devel systems so that you dont reach that limit too soon. finally, once you SOLVE the problem, THEN you know how much cpu, ram (etc) you'll finally need when the v1.0 release time comes.

to just limit your power ahead of time - that's a losing battle. sad to say, many are forced to play it that way and do extra 'spins' for no good reason at all.
 
  • Like
Reactions: insaneoctane
I'm speaking outside my expertise here, but here's my intuition about NNs and sensors:

The higher the resolution of your sensor, the more data (in terms of quantity and diversity) you need to train the NNs. Otherwise, the NN will not be able to generalize the data enough to make predictions (since the data detail is higher).

The more sensor types you use, it's the same story: more data and diversity to catch the edge cases where the sensors conflict.

By sticking only to relatively low-resolution cameras, Tesla can focus their energy on the data management side of the challenge, rather than aiming at a moving target. Also, the edge cases become more predictable and relatively easier to deal with.
 
Well, without data to show FSD is better than humans, I think getting regulatory approval will be arduous. I believe showing FSD in xx% better than humans makes this increasingly easier, because at some point it’s a no brainer…My opinion is that L5 FSD, when approved, will be at least 10x safer than human drivers.
Waymo and Cruise haven't proven that their vehicles are safer than humans and they have regulatory approval. It's not that hard to get regulatory approval.
I think 10x safer is going to be extremely difficult. I think they will try to deploy widely when they get the at fault accident rate much better than a human but the overall collision rate will be similar to human drivers. This will be due to them being less predictable to human drivers and not as good as humans at predicting the actions of bad human drivers.

"Generally, disruptive innovations were technologically straightforward, consisting of off-the-shelf components put together in a product architecture that was often simpler than prior approaches. They offered less of what customers in established markets wanted and so could rarely be initially employed there. They offered a different package of attributes valued only in emerging markets remote from, and unimportant to, the mainstream."
What do customers in established markets want that Tesla will offer less of? What attributes of Tesla approach are valued only in emerging markets? Why are those attributes unimportant to the mainstream?
 
  • Like
Reactions: Microterf
If we think of "established markets" and "mainstream" as robotaxi pilot projects, e.g. Chandler, AZ, and "emerging markets" as ADAS, then the analogy kind of works.

The analogy works better if you replace "customers" with "investors".
ADAS is in almost every new car today, you can't get much more mainstream that. Obviously going vision only (which other manufacturer are also doing) is a great innovation for bringing down the cost of ADAS.
Robotaxis are an emerging market. The advantage over the mainstream market for self driving cars is that they can support a much higher price point. They also can be much more restricted than what the mainstream market would require. Not sure if they're a good investment though...
 
View attachment 676921

"Generally, disruptive innovations were technologically straightforward, consisting of off-the-shelf components put together in a product architecture that was often simpler than prior approaches. They offered less of what customers in established markets wanted and so could rarely be initially employed there. They offered a different package of attributes valued only in emerging markets remote from, and unimportant to, the mainstream."
It’s interesting that you cite Clayton Christensen. He didn’t view Tesla as a market disruptor of the auto industry. Google it and you will see he said, “Tesla is not disrupting the auto industry, it's just making existing vehicles better, Clayton Christensen of Harvard Business School…”. I didn’t agree with him at the time. Sadly he is not here to give any further perspective.
 
It’s interesting that you cite Clayton Christensen. He didn’t view Tesla as a market disruptor of the auto industry. Google it and you will see he said, “Tesla is not disrupting the auto industry, it's just making existing vehicles better, Clayton Christensen of Harvard Business School…”. I didn’t agree with him at the time. Sadly he is not here to give any further perspective.

I wonder what Christensen would say about Apple and the cellphone industry. To substitute his thoughts as applied to Apple... "Apple is not disrupting the phone industry, it's just making existing phones better..." Just because he's from HBS doesn't mean he's right. :) I think most would agree this is an incorrect statement as Apple clearly disrupted the industry. So his statement on Tesla is equally incorrect. Tesla will do the same to legacy automakers and clearly has already disrupted the industry.
 
  • Like
Reactions: Skipdd