1) Who else thinks data is likely to be key to solving Robotaxis?
For example, Alex Krizhevsky (behind Alexnet which kicked off the deep learning revolution in 2012), quoted after working at both Google and Waymo:
For example, Alex Krizhevsky (behind Alexnet which kicked off the deep learning revolution in 2012), quoted after working at both Google and Waymo:
2) “I think Tesla has the unique advantage of being able to collect data from a very wide variety of environments because there are Tesla owners with self-driving hardware all over the world,”
3) “This is very important for machine learning algorithms to generalize. So I would guess that at least from the data side, if not the algorithmic side, Tesla might be ahead.”
4) I think the industry finds it so hard to pivot to Tesla's approach for several reasons:
A) The consensus Hardware heavy, Data Light approach originated in the DARPA 2007 challenge before huge advances in machine learning. Hence strategy was set before ML toolkit was available
A) The consensus Hardware heavy, Data Light approach originated in the DARPA 2007 challenge before huge advances in machine learning. Hence strategy was set before ML toolkit was available
5) B) As a result many self driving project teams have been led by roboticist/Lidar experts who do not have a deep understanding of ML.
6) C) Most "self driving experts" consulted by corporates, investors & the media also originated from the DARPA challenge & therefore have invested their whole careers in the Hardware Heavy, Data Light approach. This has convinced CEOs & investors etc to bet on the wrong strategy
7) D) Waymo/Cruise etc have now invested so much time and resources into their approach it is difficult to admit they were heading in the wrong direction.
8) E) Everyone else is now so far behind Tesla in their ability to role out a mass produced self driving data filtering, processing & collecting consumer fleet, now their best hope is to continue striving to solve self-driving without the data.
9) F) Many people believe 10s of millions of miles of real world driving experience is enough to train the car & to verify safety. This is possible but I agree with Tesla that it will likely require 10s of billions of miles of driving experience to meet enough varied scenarios.
10) ... Particularly as there is a fatal US accident only every 72 million miles. I find it very hard to see how you can predict & solve enough of these potential fatal accident scenarios, & then prove your software's safety, with only Waymo's cumulative 20 million miles.
11) G) Many people really believe simulation can substitute for real data and believe a software developer in silicon valley has the imagination to anticipate, model and solve all possible driving scenarios including the actions of other conscious road users.
12) H) Some people really think it's impossible to get distance & velocity measurements from cameras alone (despite the human proof of concept to the contrary & the abilities shown already even by Tesla's soon to be discontinued current 2D architecture).
@threadreaderapp unroll
