Great explanation. This is really, really complicated and I feel like most of the industry tries to gloss over how hard it is with lots of hand waving and unrealistic promises.
I love my Comma 2. I use it everyday, both for freeway and around town driving. (No, you are not my mother.) It is by no means "full self driving", but rather a tool which lets me take my hands off the wheel for most of my travels. I highly recommend Comma, even at the higher prices for Comma 3.
Great presentation. Gives a lot of detail about how autonomous systems function and the challenges. One point though. While humans have many senses, they don't have radar or lidar, and yet (when paying attention) are generally successful at driving a car. Even a car using a camera based system, will know it's direction of travel, speed, and rate of acceleration/decelleration. The FCW and AEB in my Tesla seems quite able to tell that I am not slowing quickly enough to avoid hitting the car in front of me. I think the big challenge here, is that autonomous vehicles will soon be safer than humans, but not perfect. They will have fewer accidents/injuries and deaths per unit of driving, but will also occasionally make a mistake that the average humuan would not make.
That was a really impressive video, you have such a good way of explaining things. It is a fascinating subject and I am glad you got to make content on it. 👍🏻
I would like to say, that this is the best segment you’ve done. You kept the explanations short, The humor cute and short and stayed on point. Good Job. Thumbs up sorry I can’t subscribe yet I have heavy medical bills But once I’m done with the treatments I will subscribe. TE team needs to keep going and informing their audience in the high quality style you all perform.
Thank you so much! Good luck with the medical bills - I'm sorry to hear that you're struggling. Medical bills are terrible and you should certainly not be focussing on us when you need to keep your head above water Just to clarify, subscribing is free - it just lets RU-vid know you're interested in our channel :) ^Kate
Identifying objects... I once - just once in my life - had two pidgeons flying in formation and so deep and close in front of my car from left to right, that I needed to brake to avoid a collision with the windscreen... I also honked, but, well, I don't know if the honking had any effect at all ;-) (I knew that I had a good distance behind me, so it was possible to brake without causing and accident with a car. I had that in mind. Luckily I knew the conditions around my car - situational awareness...) I hit one of them a bit, but could not see them anywhere, so they probably flew away and the windscreen survived also without a crack, but with a white powder-ish substance at the point where I hit one of them. I had just a fraction of a second, but I'm sure, I reduced the speed significantly, so it likely helped. I am sure, this is a very rare corner case, but it can happen. There was even a pidgeon being automatically photographed while flying too fast in a 30 km/h zone somewhere in europe in the last years ;-) So they sometimes fly low across streets... And they are rather big and could break a windscreen, depending on the speed difference between you and them.
Hey there, this is a fantastic video and was very insightful. I actually own the comma 2 that I have installed onto my 2019 Honda Accord. Specifically on the highway, fantastic does not even begin to do it justice and as someone who's very focused on clean, attentive driving. I find that often, I am being outdriven by the comma.
Great subject traitement , Thanks Kate , with special mention for your cold humor.
3 года назад
Comma AI just tries to keep a straight line though and now start/stop for a straight line. Something Tesla has long since solved. Comma at that very presentation said they are scared of cars that go around corners and wont do that. So right now I am not sure what kind of 'self driving' they are trying to achieve and what they want to do very long term.
Very interesting video, well done explaining all of this, I personally would love a deep dive. I haven't got FSD beta yet because I didn't even request it until a few days ago, but Tesla autopilot really needs to learn that "the speed limit signs are the speed you should going when you pass them, not a vague suggestion for the future" bit from around 8:15, hopefully FSD beta has alreadylearned that one, cause the autopilot I have now doesn't change speed until after it passes the sign, even if it's a way lower speed limit than the speed it's going, then when it does pass the sign, it slams on the regen because it's now speeding.
Great video. The bikerack example made me think about this fear of 'USA bias' that self driving AI might develop, because most of the development and testing is done in that country (and then seemingly preferably in places like Las Vegas and Phoenix.) The wide, straight roads and relative lack of pedestrians and bicycles (and bikeracks) in the USA compared to, say, Amsterdam (where I live), an Italian mountain village, let alone traffic in a city like Jakarta or Lagos, might create the sort of AI training bias we've seen before. That would be something to ask a comma.ai person about.
I have not seen a demo of any system that approaches the comprehensive approach of Tesla, and I do not think the system described is any better than Tesla.
Tesla's method is pretty lacking in sensors that will really limit how far they can go down their road. I can avoid accidents because my other sensors of hearing, feeling what the car is doing, etc. tell me much more information than just my eyes and a human brain can process the information much faster and better, as Kate described. The first 80% is relatively easy. That last 20% is a decade way at least. Especially since society first has to get past all the legal cases of determining fault in the fatal accidents and who pays. Right now my insurance does if I mow someone down but later it's all out of my control and the system manufacturer is at fault. They haven't even gotten to the unavoidable accident dilemma yet. Again, out of my control; does the car decide to kill its occupants or the people in the other car? Or maybe to protect the industry it will all become no-fault insurance, or no longer have accident insurance at all and we all just take our chances and trust in the machines (programmed by flawed humans).
I understand they might not want to use all types of sensors but 2+ makes sense. It would be good to know the density of an object, like concrete versus a cardboard box. It should see in the dark. Avoidance is best, but It should make tough decisions like - hit this car to avoid a fatal accident with another, kill 1 person instead of 2. Another level could be FSD cars communicating with each other.
Yes please! I want to know what it legitimately helps with -- does it make driving safer? Only really over long distances? Only really if you drive on big roads / highways (I live in a small town)? etc.
A coworker of mine has this running in his Diesel VW, he regularly uses it to cover the 100 miles to our company's research facility and finds it most convenient when making sandwiches on route to work.
Which is exactly the reason why using this is illegal. He should not be making sandwiches on route, he should be paying attention to the road. Even a system with over a dozen sensors requires you to keep your eyes on the road and he thinks he can trust a system that can only see a fraction of what those built-in systems can see and analyze.
@@Hans-gb4mv Thanks but I have to state that you don't know where this is being done thus have no idea of the legality within the region, and B if it's even on a public road surrounding said facility.
@@1grizzlyrizzo I'm sorry, when you say that he uses it on his way to your company's research facility, 100 miles away, it is safe to assume that it is used on public roads. You don't make a detour on a 100 mile trip just so you can make a sandwich while the car is driving because that detour is costing you time and you might just as well pull over. There aren't that many private roads in the world where you travel 100 miles between locations. And facilities that do have such roads often apply the same or even more strict rules as on public roads. You are right of course that I don't know exactly where. But unless you now say that it isn't, which you don't, it is safe to assume that at least a major part of the journey is on public roads. And I don't know of any country where you don't have to pay attention while driving your vehicle. There isn't a system in the world where the driver does not have to pay attention to the road. Even the people at comma.ai state that you have to keep paying attention.
@@Hans-gb4mv I'm sure we could gone on down the path of arguing the finer points sandwich construction and the concentration vs dexterity required, going on to examinations on transferred butter or spreads maybe even jams interfering with use of the controls when called upon to take control of the vehicle. But I will simply say this, above are the facts take them as you will, I am not my coworkers keeper and he does as he will and will not expand upon nor justify his actions.
Self Driving for me means an urban technology that may not be suitable in rural communities. I can fully support it for the mass transit and commercial users, and even for commuter traffic. Beyond that I think it is still far too early for introduction on to an unsuspecting portion of the public (the other driver).
Tesla FSD is the only viable solution when you drill down to the fundamentals of autaunomus driving. Tesla has the most experience and most data and they have deleted sensors and gone to 100% vision for good reasons. Too many seosors just crates too much input. Vision in a 3D vertorial space + time is the way to go.
@@CHIEF_420 No, multiple sensors is confusion. Redundancy is using more cameras with overlapping fields of views. If you rely on multiple sensors and lose one, you can't drive, as you are missing the data you need. If you lose one camera of many, that's no big deal, as long as you don't have big blind spots.
I always wanted to get a Tesla because of the selfdriving not because it is electric then got into comma ai back when they used Oneplus phones and anyone could load up comma.ai on their phone(Oneplus recommended). I go interested it so much that I’ll rather get a good gas car and slap the Comma 3 on it. Great video on breaking down how “autonomous” driving works.
These devices should be compulsory (when fully developed) The EU will allow Autonomous vehicles in Bus Lanes next year, there should be Autonomous vehicle only lanes on all motorways by 2025!
disagree with every fibre of my soul. I will never ride/trust or buy a self-driving vehicle. I LOVE Driving. As long as I have eyes, fingers, feet and a working brain....I want to be in control. What a horrid, boring world it would be, if we all sit in boxes to be moved around...
@@swannvictor1388 I love driving aswell just too many idiots on the road let Autonomous vehicles take them off the road then it might be safe for us (if you are a good driver)
Teslas actually do use sensor inputs from Inertial Measurement Units (IMUs) in addition to vision -- it's just not much talked about for some reason. This data is not visualizing the environment like vision and radar do -- so sensor fusion problems don't arise. And the addition of sound input (from one or more existing microphones in the cars) is also in development, per a tweet from Elon earlier this year.
This would also be a great segue into how car makers are seemingly making it hard for people to integrate their existing cameras on current vehicles two systems like comma AI. The walled garden approach to canbus and current driver assistance systems does seem to be adverse to promoting safety in general.
Several of Transport Evolved’s thorough reviews are useful to tell me what technology not to buy or use. This video, about Comma, is an example of that. Comma is another augmented cruise control system which always needs supervision by the driver (level 2). I’d rather just do the driving myself. The only current system I would trust would be the one in Phoenix, Arizona, from Waymo (level 4) which is actually self driving. Waymo will eventually expand this to parts of San Francisco.
I've been interested in comma and others that can be put into other cars. Consumer report gave it some credit when they talked about it . And I would be interested in any info or comments. Comma seems a little to diy for 3k to be spent with so much uncertain in the product.
@@GenuineFlolie One day I was riding on a city street here and was forced into a pothole almost the diameter of my bike tire and about 6-10cm deep. Broke 2 heavy duty spokes and slightly dented a velocity psycho double wall rim, needless to say I was upset.
Just part of the learning process. A camera alone see's a puddle but it could be a foot deep pothole hiding. That's the problem with vision only and even humans have to decide how to handle it. Stab the brakes and gingerly edge around it, lurch into the next lane, take your chances?
@@davidmccarthy6061 Yep and that is the crux of my question. Engineers often see the world as perfect. For instance Ford Focus Electric runs the cabin etc from a 12v sla that is charged while the charger is connected. If the person disconnects as soon as the car says it's full. And it's a cold night or someone left a usb stick in the media system the owner may start on a drive and 10min later be stuck with a dead car with 95% drive battery. (I know one of these people.)
Trying to make a mostly vision based system that works with dozens of cars, especially without tens of millions of miles of data to train your AI seems like a much steeper challenge than Tesla has just dealing with less than 10 models. Without some kind of mindblowing proprietary breakthrough it is hard to imagine how comma.ai could be anywhere close to Tesla's level of FSD.
Very interesting video. (I appreciate the objectivity this channel attempts to maintain. Far too many fanboy advocate channels out there pumping certain EV companies.)
Uses the systems for lane keep assist, lane traceing, active cruise control. and other active safety assist systems. its limited to the limits of the active systems in a car. The system uses a cameras mounted on the device. So it's limited to the front onlym So even if comma system is as smart as the Tesla, it's limited by the capability of the car. You can mount it on a Civic, but the active safety only turns the wheel partially. I've tried this it's a good system, but the physical limits of the active safety systems of your donor car really limits it to the express way. However... If your planning on spending 45,000 on a car and want to add comma ai. Just buy a Tesla SR+. My opinion.
I need to reread the Consumer Reports article, but I believe the only way they were able to justify the claim that other systems "outperformed" Tesla AP was by giving massive weight to the "safety" of the system, a.k.a. driver monitoring. Because most Telsa's don't have any driver monitoring, they penalized them. But I'm really skeptical that Comma truly outperformed highway Autopilot.
I do not think "who would win?" is the correct question to ask, it kinda implies that the first company to get it right will immediately force the others to stop. In reality, one company will get it right first, then others will catch up eventually. Who would win? Customers.
Speaking of AI and nerding out when your done subbing to and watching everything TE has done check out Robert Miles. He's a literal AI researcher who explains pretty well how hard 'teaching' AIs actually is.
You hear from real AI researchers on the subject about as often as you hear the skeptic's point of view of people claiming to be psychic. Unfortunately the hype and nonsense is more lucrative for the media.
@@2LegHumanist Don't judge before you watch his stuff. He's not in it for the money, nor is he on major media. He's been on Computerphile to explain things too, but he goes more in depth on his own channel. At least watch his "Intro to AI Safety, Remastered" video, it's only 18 minutes long.
I don't see the need to identify what an object is as long as the AI knows to avoid it. Identify it's size, its bounding box, assess it's speed and direction, and if you're on a collision course with it. Who cares whether it's a tandem bike or a rhinocerous? *Just don't hit it!* By the way, fond memories of SiliconGraphics and that great logo. Didn't they use that Alpha chip or something?
Everything comma is trying to do,Tesla has done some and is tackling some right now,also Tesla FSD Vision is currently driving much more better without those sensors.
Very interesting and informative video, it shows an almost infinite number of variables the system must evaluate and act upon. True autonomous driving capability equaling that of a human will eventually be achieved but probably is decades in the future. In the mean time driver aids that take some of the monotony out of daily driving are here now and will continue to improve.
Why does the AI even need to classify the object? You're not supposed to hit /anything/, no matter whether it's a car, a bike, a pedestrian or a lamppost. If it doesn't look like a piece of road, don't try to drive on it.
Because depending on the type of object, the AI is going to predict different behaviors. Static objects are relatively easy, but LOTS of things in the environment an autonomous vehicle has to travel through move, and move in very different ways depending on what they are - Winter
They did mention that comma.ai does not provide equipment for reviews, and also that it would cost around $2000 to purchase one. Wow i was not expecting that price! Granted, compared to the price for Teslas assistance package that is a bargain!
@@useaol yes the comma 2 is cheaper but the comma 3 is much nicer and what you really want. There’s a pretty big community. Surely a nearby ride along could have been arranged. I dunno. Fair enough.
It's extremely difficult for a human to remain fully engaged with a vehicle's driving situation if the AI is driving *most* of the time. Thus, subsequently "handing off" vehicle control to a human driver in the midst of a confusing situation, is a recipe for disaster.
