This means that anyone doing self-driving needs to be a manufacturer in order to avoid getting squeezed on terms and profitability. And vice versa. A partnership won’t cut it; one company of the partnership will get screwed and the thing will go nowhere.
That’s why having both is critical, and I think this point is not fullt appreaciated.
Waymo is the strong party in a negotation with any manufacturer, and the manufacturers know this.
If Tesla has every one of their cars feeding their models more data, it does feel like a natural conclusion that they will be able to overtake waymo.
*Disclaimer: I'm not an expert in ML
Tesla has built their own chip optimized fully for self-driving.
all-in on vision based full self-driving, which seems to be the only way to solve all the edge cases. Even if google accomplishes self driving with lidar, lidar will never be able to compensate for all edge cases. Human vision already compensates for all edge cases, thus the case for vision based self driving.
Another way of asking this question: how many "millions" of miles is sufficient, and at what point are there diminishing returns?
The problem is slogging threw millions and millions of corner cases in the real world. A trillion miles on the highway are quite pointless, a trillion miles in one Arizona city will also not get you there.
You need millions of miles in all driving conditions and train the car to always figure out 100s of details. Look at the Tesla presentation where just recognizing a 'Stop' is incredibly difficult. Working threw all these problems requires having 10000s of cars with real sensors sourcing data from you.
The advantage Tesla has is that they can do this with zero capital, while Waymo needs to have a expensive cars (100 to 1000s) with a payed driver, go threw every location and create a HD maps.
For starters, Tesla still seems to have this annoying habit of not detecting large, white trucks that are stopped in the middle of the highway, and will happily drive into them at full speed. This isn't exactly an "edge case" issue here.
The "real-world driving data" is what I'm calling the "raw data" here, as it's the raw data from the standpoint of the (perceived) AI algorithms driving self-driving cars.
> Tesla is the only one going all-in on vision based self driving. That's the key thing that's going to get us to full self driving.
Why is it the key thing? Why should worsening the quality of your input data improve the results?
Our eyes get worse data than lidar, yet our brains are able to process this data more effectively.
Also, lidar is super expensive. It will take much longer to get affordable lidar into everyday vehicles.
Running with a more responsible hardware stack means the software has to be more cleverly designed, and that design is mostly not in the just-throw-data-at-a-neural-net stage. Building your visible-light-only driving system is going to cause you to struggle a lot harder at getting a good subroutine for "how far away is that object?" while providing you no benefit to solving the harder problem of "what will that object react to my coming nearer to it?"
<joke>To be fair, it was an edge case until Tesla started selling cars.</joke>