So Elon is ~100% confident on his approach towards autonomy but why is Google/Waymo following a different approach? The answer, I think, lies in the genesis of their efforts towards autonomy.
Elon's approach:
- Elon said he felt he needed to pursue autonomy because in his mind non-autonomous vehicle would be less valuable than autonomous vehicle in the future. So not to pursue it may even defeat his primary mission of electrification.
- Once he realized the need to pursue it, as usual, he followed first principles thinking.
- Broke down the problem and started solving for each individual lego piece the best way possible -
- Collecting huge data (the very basic requirement for NN training) by putting in required hardware from the very start on every vehicle even if the customer decides to not pay for AP/FSD,
- Developing a very specialized NN chip for over 3 years,
- Hiring the best AI researchers to build the algos (probably the easiest decision).
Google's approach:
- On the other hand, I think Google went into autonomous vehicles primarily because of one of their expertise - MAPS.
- Google probably thought because they have spent so much of their resources on building out the maps of the entire world - what next, what other ways to montize this expertise?
- Answer - build even higher resolution maps and leverage their own AI researchers (also best in the industry) to work on rest of the stuff to pursue autonomy.
- I think they will achive autonomy but it will be stuck in Geofencing and it's just not easily scalable.
- Also not to mention the LIDAR approach which will help them to achieve local maxima, but still really have to solve vision to really grind towards global maxima.
- I think the partial reason to select LIDAR was because they think solving vision was very difficult - primarily becuase of unavailability of huge amounts of data (including weird edge cases that simulation just can't predict and only real world data can provide).
I am just guessing if Google had started on automomy from a blank sheet of paper, they would have followed a similar approach as Elon's (maybe by partnering up with an auto manufacturer to setup the hardware to collect the real world data).