Some interesting comments from that YouTube thread (following the video):
Dan Fitzgerald
2 hours ago (edited)
As a robotics engineer, I see two very different technology stratagies. You can either focus on really good performance in a limited area, then expand that area to scale (what everyone else is doing), or focus on decent performance in any area, then improve that performance everywhere (what Tesla is doing.) Understanding this difference is critical to understanding why all the other L4/5 efforts are failing. They look great in their geofenced areas with custom high-res maps, but can't generalize. They didn't anticipate the "long-tail" of challenging circumstances encountered in real diving.
Chris Wright
1 hour ago
I agree, but I would lump Tesla in with all the other failing systems. The technical challenge of true self driving has been vastly underestimated. I was a little surprised by this CNBC video representing the failures as business decisions rather than being caused by realizations of the technical challenges ahead. I really believe some kind of artificial intelligence is needed for level 4 and above (as opposed to 'automation' or, as I think of it, trying to write code to cover all eventualities and scenarios....to bludgeon the system into working through coding). An ability to solve novel problems (something humans are actually very good at) is required. Tesla is not even close to this.
Hans
47 minutes ago
@Anirudh Kamineni I feel you, and Dan, are getting a bit confused. Yes, 4 > 2. That's a fact, both in mathematics as well as in autonomous driving. If you don't recognize that, you don't know what the levels mean. In L2, you as a driver, have to keep your hands on the wheel and your eyes on the road. With L4, you are no longer required to do that. The big difference between L4 and L5, and why Dan is already a step behind in his write-up is that L4 is always geofenced while L5 isn't. The number says nothing about the technology used, but everything about how autonomous the system really is.
Tesla is aiming to go from L2, where it currently sits, straight to L5. Mercedes and others are going through all the stages, without skipping one, building up that way.
What Mercedes did is a first step. It has shown that it can handle L3, both from a technological point of view as well as a regulatory point of view. They can now start building out. Go for higher speeds and bigger areas.
And is Tesla really that different? How did their FSD start back in the day? Wasn't it similar? You start on the highway, with navigate on autopilot. Then you slowly start building out. And that is also what Tesla is doing to this day. The current FSD beta is only designed for North American roads. It's not as if Tesla can go for full regulatory approval in the US and Canada and then say: now we flip a switch and let the rest of the world join in. They just can't. Once North America has been done, they have years more work to do the rest of the world.
And in all honesty, I doubt we will see regulatory approval for Tesla's FSD within the EU before the decade is out. I'm still waiting on that coast to coast trip Elon promised in 2017
Dan Fitzgerald
46 minutes ago
There are at least four other ways that Tesla's strategy is different from everyone else's (to my knowledge).
1st: they design their own autonomy computers and components instead of buying from third parties and suppliers (much more vertical integration) and also design their cars from the ground up with the autonomy stack (no mounting sensor boxes on existing non-av cars.) They also have much greater control of the system integration and coordination between the autonomy team and designers - the sensors are not an afterthought and the teams are no siloed.
2nd: they rely almost exclusively on vision, not lidar. Cameras are cheaper than any other sensor.
3rd: they learn from their fleet of human drivers (which is larger than all the others combined) 4th: they are pushing for end-to-end deep learning, and do all of that infrastructure in-house (dojo).
In my opinion, all of these are the right way to go, and any one of them alone would confer an nearly insurmountable advantage over competitors that are not doing it.
In tha 360. MRF User Map
