Not buying the storyline. Certainly Tesla will have to keep upgrading its systems to achieve true FSD. They could even decide that LIDAR has become cheap enough to use anyway. What they have is cars on the road collecting data. They'll have to collect new data and develop new algorithms on that data but that path is not excluded from them.
> "It is a misconception that you can just keep developing a driver assistance system until one day you can magically leap to a fully autonomous driving system," Krafcik said. "In terms of robustness and accuracy, for example, our sensors are orders of magnitude better than what we see on the road from other manufacturers."
Not a good argument, very handwavy. Better sensors is not a moat. What is your software, how is it qualitatively superior? No one suspected that NNs or even DNNs could produce AlphaGo/AlphaZero until it did.
Until Waymo gives their tech to the public and we can compare how well their less used but superior hardware and software performs it's all opinions and PR.
The proposition that is being discussed is not whether Tesla as a company will get to autonomous driving, it’s whether Tesla’s current suite of sensors and its “HW 3” inference engine will get there. Keep in mind, Tesla has taken thousands of dollars from customers with the promise that they will deliver them “full self driving” in the near future (or actually in the past, as it was promised by 2018).
If Tesla revises their Sensor package to include LIDAR, that would be a sign that the “Tesla approach” has failed.
I agree. The entire Tesla gamble really hinges on whether the HW3 + current sensor suite is sufficient for them to achieve a safe enough degree of autonomy. It's about as minimal as it gets from the hardware front, and is entirely dependent on the machine learning side to evolve rapidly along with rapid software update deployments.
I feels there's a broad misconception across many of these types of discussions, where people assume that just because Tesla has not chosen to widely deploy LIDAR across their vehicle fleet, that they do not understand it's benefits nor value them. This is entirely false. For several years now, Tesla has driven test vehicles with very high end Lidar units to basically generate "ground truth" understanding of the world that can be then used to test and improve their camera-only FSD inference. They heavily relied upon it to build ML models to generate depth maps across frames, and that capability is already widely deployed via Autopilot (poor man's FSD). That it works in itself is somewhat shocking to me.
Lastly, if there's any indication of a mea culpa coming from Musk on this front, those ideas should be easily put to rest. Recent FSD footage shows rapid and drastic improvements to the capabilities of the FSD-beta system, and Musk himself is still repeatedly bellowing about how his personal vehicle with the latest software is able to handle more and more challenging scenarios. If anything, Tesla is doubling down, and expects to win. Given their history, if I was a competitor, I would be troubled at least.
Why? You can argue that CV is enough for SDC AND add extra LIDAR sensors to make it even better/safer (esp if the cost curve of the sensors make sense for consumers).
Just because I can ride a bike doesn't mean I shouldn't wear a helmet
> Until Waymo gives their tech to the public and we can compare how well their less used but superior hardware and software performs it's all opinions and PR.
What constitutes giving tech to the public? They are in Phoenix and they say they will be rolling out in other cities soon. So you can go to Phoenix and hail an autonomous ride to compare yourself, where you will find that no else is even close to offering truly driverless rides like Waymo is.
Waymo is amazing, but it doesn't make the business profitable: it needs to scale up the investment in hardware and keep its advantage until Tesla gets good enough to compete. This PR is mostly aimed at investors in Waymo.
"What they have is cars on the road collecting data" ... what do the ToS say about this? Is this really true they are gathering data from all their customers? I havent seen uptick in my internet bandwidth as Tesla owner.
If you are using AP or FSD and you take control suddenly, that will sometimes send data to Tesla so that they can improve the autonomy. I don't know what the ToS say, but Tesla are pretty open about the fact that the cars do that.
I expect Tesla and Waymo to be the 2 best driverless solutions in the near future. Tesla is a real competitor to Waymo, as the better sensor has a litmited time advantage until deep learning algorithms improve even more.
Waymo doesn't have too much time to get back the billions of dollars of invested R&D money, while Tesla is already profitable even without self driving.
I've never been convinced that deep learning is the correct approach for something so nuanced as driving a car. Deep learning is excellent for generalisations and consolidating data into single metrics. It's not so good at identifying edge-cases and reacting accordingly, those cases are more likely to be merged into other reaction types.
Ha. Have you ever tried to play Go? When you feel competent, learn that there are still many orders of magnitude that can't even be told because the language is completely foreign to one who doesn't already think in it.
> "It is a misconception that you can just keep developing a driver assistance system until one day you can magically leap to a fully autonomous driving system," Krafcik said. "In terms of robustness and accuracy, for example, our sensors are orders of magnitude better than what we see on the road from other manufacturers."
Not a good argument, very handwavy. Better sensors is not a moat. What is your software, how is it qualitatively superior? No one suspected that NNs or even DNNs could produce AlphaGo/AlphaZero until it did.
Until Waymo gives their tech to the public and we can compare how well their less used but superior hardware and software performs it's all opinions and PR.