Reciprocity
Active Member
Yeah I am skeptical that Tesla has autonomy magic. I am sure their CEO is putting the most pressure on them however, and that counts for something. Tesla doesn't really need to necessarily be the tops forever. Its not really a critical thing. There ARE points for second place. If BMW gets to 99.99 while Tesla is only 99.9 and then Tesla gets their a year later, that is not an existential crisis, IMO.
I also wonder about the "fleet learning data" that is such a religion around here. I certainly hope and pray they have some great tech that allows ghost data to form autonomy... it's theoretically possible. It just sounds like a seriously hard problem.
No one is even proclaiming full autonomy as a realistic idea until 2021 or later. Only Elon has made the bold proclamation that it is he considers it solved and they only need time to complete the machine learning software. Only Elon says you will be able to go coast to coast and parking lot to parking lot by Dec. Only Elon says you will be able to feel comfortable sleeping while your car drives you to a destination in 2 years, which would be mid 2019. I agree that Tesla will release something before anyone else and it will be 99.9% and will save half the lives of a system that is 99.99% but that will be good enough. Now the question that I have is whether that will be in 1 year or 5. I believe they have chosen the right path, meaning GPS + Vision + Radar and no Lidar. I dont think its anymore complicated to do this without Lidar then it would be to do it with a Lidar, that btw does not exist yet. Meaning there isn't a lidar you could put in a car that you could sell a half a million of, the costs would be to high and the lidar not good enough. I am certain that Waymo and the others will get lidar to work and will shrink it down and cut the costs, but it wont happen before Tesla has a half million cars on the road.
The data advantage is real. Its real because how machines learn. We learn by having an instructor show you how to do it. Machines learn by millions of iterations where they can see a picture of a stop sign from every angle and every situation, snow, rain, dusk, dawn.. You and I can see a stop sign and our parents can say.. that means stop. And we can recognize that sign in the future. Machines need to see millions of examples and not just signs, but the context with which they are in so that it can ignore a stop sign on someones t-shirt or a stop sign that is tiny in a picture on a billboard. This is impossible to do if you do not have billions of miles of real world data. You cannot find enough pictures of stop signs in any database. That is one tiny example of why the data is so critical and why its harder for others to get. Mobileye has some of that, but they dont have a currently large volume of 8+ camera cars, only a ton of single, front facing camera cars, like AP1.
Last. BMW wont be developing this, they will have to buy something that has been developed like GM and Cruise automation. They are not a machine learning tech company. They dont have 50 engineers that are focused on deep learning neural nets, vision systems and machine learning AIs. They are an old world company and they will move very carefully and slowly.