Did I miss something? When did camera only autonomy happen?
Give me a break. James Douma is a Tesla youtuber that has zero items on his resume that makes him an expert on anything except PhoneGap. AK hasn't said anything about a date or "soon". He has said vague things like "i think it's possible at some point" (which may be in 100 years). Hit me with that Weinberger quote please.
There are 40-50 years of research that I can point to. Automation complacency is a real thing. Here is one random paper: Complacency and bias in human use of automation: an attentional integration - PubMed
ME says they will not do camera-only for AV, because it's not safe enough for autonomy. Autonomous isn't "can drive". It is "drives safely and 'never' crashes".
Oops. I thought this was one of the “every expert doubted Tesla and look at how wrong they were” posts.
They appear to be using HD map data. The visualization is clearly rendering sections of the road that are not in view of the car. Tesla's feat is that they're using vision not only for object recognition, but also lanes and road edges.
spectrum.ieee.org
No. English is not my first language, but "coming years" to me is at most 3-5 years. Again, let me know when you find a single scientist that says camera only autonomy will likely happen before 2030, or even commit to a date. You won't even find a Tesla employee or a person with vested interest such as AK saying such things. How come do you reckon?
All he says is that it's Teslas big bet... Why are you implying that he says its solvable? in the coming years?Tesla Places Big Bet on Vision-Only Self-Driving
Full Self-Driving beta software v9.0 shows many improvements, but the road to full autonomy still looks rocky
Again, I linked to a random article This is a well-researched area. Do you dispute that automation complacency is a thing? If not, its easy to understand that the system is the performance bottle neck. And if the system is less safe than a human, then that's how safe the combination is when the system is good enough and complacency sets in.
I put 'never' in quotes, as was obviously talking about a system that has superhuman capacity or at least very high MTBF, which I am sure you understood and you just happened to remove when quoting me...
Nice summary. Assuming you're talking about L5 and before 2030, I am definitely in the the "not possible" camp. L4 multi-sensory is already here. and L3 multi-sensory too in trivial ODD:s at low speed.
In my view, the largest problem is that people tend to get their "education" from the likes of James Douma and Dr Know-it-all. The power of Tesla marketing...There are also 3 types of TMC users:
1) Uneducated/Uninformed - these are the people who have joined the Tesla car family without any understanding of AI/Autonomy and saw a video or read an article about how Tesla is attempting ADAS functions. They don't know how/why it works and are reading articles/debates here on TMC and forming an opinion based on new information from their driving experience and the various information they glean from TMC.
2) On The Fence - these are people who have read and understand the various facets of ADAS and autonomy, but do not subscribe to one particular viewpoint or the other, or they are leaning towards a camp but have not fully embraced it. These users can be swayed by logical arguments made here and in various videos / articles to join one camp or another. Their opinions evolve based on the merit of a particular argument.
3) Entrenched - these are the users who have done the research, watched the expert videos, read the compelling arguments here on TMC, and have joined their camp. They firmly believe their camp is correct and no amount of argument, logical or otherwise, will dissuade them from their opinion.
All the debate on TMC is fun to watch, and I hope that people in the first two types can use the discussions to help form their opinion and educate themselves on the complexities of the world they've just entered. That's the goal - help the #1 and #2 type people.
It's human nature to attempt to persuade others to your thinking, but the issue is when two entrenched people assume the other is on the fence and attempt to evolve their opinion. It's futile - but definitely fun to watch as a spectator.
and the second largest is people that join a forum with a primary focus on proving they know everything with no contribution. Almost like they have a second agenda?
Yes. Helping people not falling for the same scam as I did, and sharing the knowledge I have gotten over the last three years? Sorry if I upset some people here with facts.
Not sure how relevant this but at least another data point as you folks battle this out.
I think everyone is going camera only for driver assist systems. The term "self-driving" now encompasses both driver assist and driverless systems. Nobody other than Tesla is working on a camera only driverless system (or L3).
But we don't drive with eyes alone. The majority of it sure. But we also use sense of inertial motion, 360° hearing, vibration (approach of trucks & emergency vehicle growlers), memory & experience (puddle over a known pothole, ice formed on this corner because it always does) to round out the 100%. Allows us to sense non-visible hazards. Also advanced visual clues like reflection, human gestures ("careful, something's around the corner"), inference (ball rolling, human likely to follow). AI may require more sensors than just cameras and these reactions may be challenging to replicate with a NN.
Yeah, and we also have a brain. A NN AI is not a brain. It doesn't learn on the job. It's a fancy statistical model.
Perhaps they can pair the NN with a local memory cache of 'stuff learned that's useful but not necessary to encode in the global model'. Like: there's a guy on this corner on Tuesdays that always throws oranges at the car, avoid driving here then. Or you know, something more serious.
The solution you are suggestion is one of many things humans can do, and a self driving system cannot (yet, who knows in 50 years right?). The general problem here is how the auto industry works with type approval, testing and manufacturer liability. They need to test the system before they ship it (supposedly, questionable in Tesla's case where they slap on a beta label like driving is like email), and guarantee its function. Can't have as many systems as there are cars, obviously. This is why on the job learning is hard for cars. Plus the science.
It’s impossible not to use maps; the question is really how much and how they will be used. If you consider you a human drives, they have some sort of map in their head to know where they’re going. Similarly, the car needs to have an idea where it’s going, otherwise FSD would just be a random turn generator. The difference is how much it can interpret, adjust and accommodate ‘on the ground’ differences from the map data. One approach is to have hyper accurate data so the car never has to interpret anything. Since roads and conditions are constantly changing this is a fools errand, IMO. I think Tesla’s approach of being able to adjust based on visual data is likely better, just more difficult to implement.
