… this actually lends credence to Elon Musk’s, “cameras only” FSD approach.
from WSJ
{{ Monday, self-driving technology company Mobileye announced it was ending the internal development of “next-generation frequency modulated continuous wave (FMCW) lidars for use in autonomous and highly automated driving systems.” It added: “We now believe that the availability of next-generation FMCW lidar is less essential to our road map for eyes-off systems.” }}
Apparently LIDAR and cameras often produce conflicting information, but the camera is more often correct – just not “100% correct”.
Less absolutely necessary? The WSJ/Mobileye could use a better copy editor.
This is what Elon Musk said a long time ago.
o o o o o o o o o o o o o o
A fellow at SA got me to read The Computer and the Brain by John von Neumann, 1958. An excerpt [emphasis added]:
It should also be noted that the message-system used in the nervous system, as described above, is of an essentially statistical character. In other words, what matters are not the precise positions of definite markers, digits, but the statistical characteristics of their occurrence, i.e. frequencies of periodic or nearly periodic pulse-trains, etc.
“Statistical” means not perfectly accurate but close enough.
Does it? From their press release, it doesn’t appear that they’re doing this to move to cameras only. Instead, they’re moving to third-party LIDAR systems and emphasizing development of their in-house imaging radar - keeping both in combination with vision:
As part of our regular review of the long-term technology roadmap, we now believe that the availability of next-generation FMCW lidar is less essential to our roadmap for eyes-off systems. This decision was based on a variety of factors, including substantial progress on our EyeQ6-based computer vision perception, increased clarity on the performance of our internally developed imaging radar, and continued better-than-expected cost reductions in third-party time-of-flight lidar units.
Seems like they’re not moving to cameras only. Rather, it’s just that they’re going to buy off-the-shelf “good enough” LIDAR systems and work on improving their radar - but still using all three sensor systems.
50+ startups with more than $1B capitalization that used Lidar based tech have already folded. Mobile Eye appears to be moving away from Lidars as well. Waymo is heavily Lidar dependent and it works well. However, it is geo fenced and it is really hard to tell how much of it is “remotely driven”.
Elon was and is right. Lidar is expensive and gives conflicting signals to Vision.
I agree with @albaby. That does not appear to be the case. Mobileye has previously partnered with Luminar Technologies to supply LIDAR. The article says that Mobileye won’t be developing their own LIDAR and goes onto say that instead Mobileye will simply be using commercially available LIDAR because what is available off-the-shelf is good enough for them.
I haven’t been following this topic too closely, but as far as I know, all the major automakers besides Tesla use LIDAR and/or RADAR. The last few months I’ve been seeing a number of driver-assist comparisons come across my news feed. Obviously, there is a certain amount of subjectivity in these, but I don’t recall seeing Tesla with the top ranked system. At least not in the past year or so.
Depends on who you talk to. Autonomous driving is still in its infancy, sort of like the Wright brothers flying machine. It is also in transition from purpose written code to neural network based AI which is a huge leap. Some facts are well established, the more powerful the data center and the more good data you have to feed it the better the results get. This correlation was highlighted by Andrey Karpathy, the former head of AI at Tesla. Several million Teslas on the road gathering data and Tesla’s huge data centers tell me that Tesla’s FSD is the horse to bet on.
Which automaker have you seen with the top ranked system? I’m not referring to the standard enhanced cruise control system (lane keeping, speed and distance keeping), pretty much all mid-to-high end cars have that already and they all work very similarly. I’m referring to the system that you enter a destination and the car attempts to drive there, doing the turns, stopping at stoplights, entering and exiting the highway, etc.
That list includes some car manufacturers and some non car manufacturers. They’re also not always compares apples to apples, the systems they are comparing are often vastly different, with different capabilities. If they want a true comparison, they need to compare only those aspects of similarity. For example, if they are comparing only highway driving assist features, they need to compare those parts only. And I’ve tried those features on 4 or 5 different manufacturers and they are all very similar (back when I did the comparison, I found Volvo to do the best job overall, but I had no access to a Tesla back then).
I haven’t done a rigorous study or anything, just noting what pops up in my news feed occasionally. But for example, Edmonds ranked GM’s Supercruise as their favorite. These guys thought it was a tie between Supercruise and Ford’s Bluecruise, etc.
You mean the one that drove over a fallen pedestrian in SF and dragged her 20 feet down the street? And then the company sort of tried to lie about what happened…then got kicked out of SF? And then it turns out they have 1.5 support and remote drivers for every car in operation?
No. That’s Cruise the robotaxi company. That’s something completely different.
The article was talking about GM’s Level 2 driving assistant called Supercruise which is roughly comparable to Tesla’s FSD.
It was claimed up above that self-driving companies are moving away from Lidar. My point is simply that the top rated systems all use Lidar and/or radar.
This is absolutely not correct. GM Supercruise is roughly comparable to Tesla Autopilot (comes included with all Tesla vehicles at no extra fee), it is not at all comparable with Tesla FSD. Both (“Supercruise” and “Autopilot”) are just fancy names for the combination of lane keeping, cruise control, and maintaining a minimum distance to vehicle in front. Almost all mid to high end cars have that feature. And the GM Supercruise is indeed a good implementation of that feature.
Tesla FSD is a MUCH MORE advanced feature. Tesla FSD costs $8k or $100/month. Tesla FSD is the feature that you route to a specific location, then engage FSD, and the car drives you there by itself (including turning, stopping at stop signs and traffic lights, entering and exiting highways, etc.) Sometimes FSD will even enter a destination parking lot, look for a spot, and park in it!
Part of the problem is the way Tesla brands their features. They often use autopilot as a superset of “autopilot the feature” and FSD. But the features are very different and are totally distinct.
