i am stading right now and giving a full minute clap of hands to you guys!! amazing!

I am really impressed by the left turn at 6:17. It starts accelerating and entering the intersection, even as the car on the right is also entering the intersection, because it can plan to go into that lane and turn just after the other car, without major risk. Really amazing.

@djmips

9 ай бұрын

and it appears that the motorcycle doesn't even stop for the stop sign!

Very educational. Thanks for posting.

The visuals of all the information the car can see is amazing!

@felipeamper7248

3 жыл бұрын

Imagine millions of cars scanning cities all the time! We would have a real time map of everything in the city (like street view but live).

I love the seamless driver assist feature.

This was very informative, Thanks.

Super impressive, also quite entertaining. Well done!

Do you ever simultaneously use drone camera vision as a secondary source for training data? I could see a drone hover-following 100 feet above the car to help validate the lidar generated map. Would something like that be useful?

@DHT2023

3 жыл бұрын

I’m wondering about the birds eye view too. How is visualization for live traffic able to see that far in all directions ??

@catbert7

3 жыл бұрын

@@DHT2023 The lidar is on top of the car.

@cybertrk

3 жыл бұрын

Lidar is this... 100ft above yields no extra data. At most you'd want a lidar 13.5 above the ground which is max vehicle height.

@catbert7

3 жыл бұрын

@cdb That is incorrect. e.g. Waymo's mid-range lidar has a range of 75 meters. Even their short-range system has a range of 20m.

That construction zone maneuver @3:12 is jaw-dropping. I literally said, "WHAAAAT...." out loud.

@unknownbrokie

2 жыл бұрын

it needed remote assistance to complete that scene

What a great conversation! I’m so excited about all the progress you’re making at Cruise and I hope you do more videos like this. The map showing everything the car recognizes and how far it can identify objects down side streets shows just how amazing the Cruise technology is. With an aging and increasingly distracted population, I really hope this will help make travel safer. I’m curious to know how long you think human attendants will be necessary to interpret things like construction zones, bad actors who try to stop the AV to hurt/rob passengers, and other totally off the wall situations.

What Lidar sensor are you using for this system, how many sensors, and what are the placement? From my experience with Velodyne, not even the 64E is capable of this amount of detection. Without having at least 4 to 6 VLP 16 sensors in a SVS configuration, I do not see this performance during real life use-cases with out post-test manipulation.

@2:55 how was your left turn maneuver performed without knowing for sure the opposing car was turning right and not continuing straight? The turn signal was barely picked up in the camera and Lidar is not capable of accurately determining turn signal intensities. Do you use a GNSS to determine vehicle positions on the roads, and with that being said, how can you be sure the vehicle is in the corresponding lane?

@FrankGonzalezIV

3 жыл бұрын

Good question!

This is really interesting. Just before 6 minutes the question is raised, "How long does it take a human to learn to navigate (do this)?" The next question would then be what learned experience do they gain and what does that look like? And while they're driving what information do they use? Experience seems well collected in this method - predicting very quickly and accurately that the approaching car was turning right. How generally would it be able to apply this learning? (As a driver with a lot of experience I've had a lot of moments where 'you just know' someone is going to do 'something', so you back off, speed up, stop following - do something so if that other driver does what you predict, you'll be safe. Of course you don't do over react to a 'maybe', and there are times when you don't do enough: slowing down, changing lanes still isn't enough when that guy on meth changes 4 lanes right, then 4 left hits two cars and suddenly 6 cars in front of you are spinning out of control) Speed. Neat trick speeding up the video 5x. It masks the slower speed the test vehicle is actually driving at. But back to what type of information does a human driver use. Primarily visual. Is that enough? I think the correct answer is, Well, it's mostly worked so far. So are all the accidents, wrong turns, near misses, out of sync with other drivers (driving too slow or too fast compared with the general safe practice- most of us think of as being the speed limit +5-10 mph), What lack of information causes these issues? And how do we get it? The proximity rear end collision systems seem to me so far to be the most significant step. When things come straight at us, or we at them - depth perception is hampered and that's how rear enders happen. All predators on their final approach go straight in, not simply because it's the shortest path, but it gives the prey the least amount of depth information. By the time the size of the attacker changes, it's too late. When we stop short , it's because we should've been slowing down, but that vehicle in front of us hasn't moved left or right, or significantly changed size - until we're closing in, then we hit the brakes, occasionally too late. Just thinking out loud.

where is the intro music from

Very impressive. Love the conservation. Wonder if OpenAI and Cruise could partner and create a neuronetwork for their fleets.

Amazing!

It pumps my heart going.

Very impressive!

Very nice video! Is there any way to interview you, Kyle? Would love to have you on my podcast!

Are any datasets like vision, lidar, radar, IMU going to be released publicly? It would be an enormous asset (even unlabelled) for academic research and open source development in general.

At situations like at 3:12.. I know a human obviously intervened. You can have an infinite number of these construction scenarios, multiply that with several workmen giving you conflicting directions or hand gestures.. and add in a temporary stoplight, with temporary stop signs (not in the octagonal shape). That's a serious challenge for ANY autonomous car!

@raygordon3728

2 жыл бұрын

That's something I don't think they'll ever solve. What about a 4 way traffic light intersection with the lights out? And a cop directing traffic? omg, that would be ugly.

Very impressive. There are some naysayers, but I feel this might be a really successful system.

You can see how they're stuck in a local maximum and need remote assistance for any situation outside the norm. They don't have enough mileage data to train their AI to UNDERSTAND those situations.

@CookiePepper

3 жыл бұрын

Agree. "75-minute fully autonomous drive" is false claim.

@qr-ec8vd

3 жыл бұрын

@@CookiePepper lol try doing half of this with tesla FSD

@zaydevans2077

2 жыл бұрын

@@qr-ec8vd the same thing could be said the other way around. Try taking this car decked out with this tech down Lombard st etc. or anywhere that hasn’t been meticulously mapped out etc

Damn this is so motivating especially your conversation at the end

super cool and impressive!

Seriously impressed with the performance of Cruise and with the genius of these CEO's

Beautiful technology I’m eager to add this to my everyday life.

What do you make of the valuation difference between GM and Tesla? Is it justified?

Wow this aged very well.

6:59 Hi Waymo :D (mapping vehicle I guess?)

That was really impressive, aside from the construction site at the beginning, 0 interventions in a busy city with lots of unique scenarios

@jamesstevens2145

3 жыл бұрын

It wasn't "zero interventions" Kyle admits, in the video, that a remote driver took over at one point

@dmitchellhomes

3 жыл бұрын

@@jamesstevens2145 The remote driver just initiated the car to continue, the remote driver did not help control the car.

@jamesstevens2145

3 жыл бұрын

@@dmitchellhomes Human intervention is human intervention. If the journey could not have been completed without human assistance then its not "75 mins of autonomous driving" - its "75 mins of human-assisted autonomous driving

@legacytesla

3 жыл бұрын

@@dmitchellhomes that's not what was said. He said the remote assistant plotted a course around. It was not simply go ahead.

Kyle, while this is undoubtedly really impressive, do you think its fair to call it "75 Minutes of Autonomous Driving", when you admit that at least once a remote driver had to take over? I think I saw a remote driver take over more than once, but I need to have a closer look

@movax20h

3 жыл бұрын

The "STATE" remained Autonomous, so it is hard to say. Also the planned route was still being displayed, so I believe it was actually Autonomous and not remote driving.

@ai_and_chill

3 жыл бұрын

it's a remote operator that offers path guidance, not a remote driver. he mentions that they plot a path for the av to take autonomously.

@jamesstevens2145

3 жыл бұрын

@@movax20h 1) I don't think the incident talked about was the only one 2) human intervention is human intervention. If the vehicle couldn't complete its journey without human assistance, then I think the description is misleading. The description of L4 is pretty clear on this - www.aiexclamation.com/levels-of-autonomy/

@ai_and_chill

3 жыл бұрын

So you would prefer a self driving taxi service that's completely unsupervised in an urban environment with an incalculable amount of variables?

@jamesstevens2145

3 жыл бұрын

@@ai_and_chill I would prefer an accurate description of the video, that's all I was talking about. But since you specifically ask, Yes L5. No L4 operator will remain in business once anybody has L5, the relative cost different will drive all L4 operators out of business. This is why google are selling off chunks of Waymo. They know its not a long term business model. They miscalculated how soon L5 would be possible, so went for an L4 solution that has no path to L5, ditto Cruise. The controls of a car are very simple. Faster, slower/stop, left, right, indicate, back-up (very rare). That's about it. Once you can accurately label the objects you see, actually driving is pretty much done. "Driving" as a complete concept is complex, but if you break it into its constituent parts, its solvable.

Smart stuff

It's astounding that this video of one of the premiere autonomous driving systems has only had 55k views after a year. I guess if people can't buy stock directly in your company, they couldn't give a flying funk about how good your system is.

I think one of the most amazing things is how it knows what a street vs. a sidewalk or other surface is. Is this from the use of mapped roads?

@CookiePepper

3 жыл бұрын

Those static objects are in the HD map.

Does it predict dangerous door opening from cues like recently parked and lights going out? and does it see a door opening very early? I suppose use of lidar makes that easier.

@Karthik-lq4gn

3 жыл бұрын

At some point, you won't need LIDAR. Deep learning systems will become so good at making 3d models of the world using real time video frames that LIDAR will only be used for redundancy (when it gets too dark or foggy).

@DanFrederiksen

3 жыл бұрын

@@Karthik-lq4gn Sure, lidar just allows for easier algorithms to avoid collisions. What lidar could be used for though is training the vision system to have good distance data and never missing anything, even though the final vehicles wont have lidar. Because machine learning paradigms aren't yet as strong as humans, it benefits from getting stuff spelled out and lidar is good for that, I believe.

@stinger15au

3 жыл бұрын

@@DanFrederiksen (when it gets too dark or foggy). Well that's pointless since Lidar is actually terrible in bad weather, no better in fog than a humans eyes and probably worse.

@DanFrederiksen

3 жыл бұрын

@@stinger15au a more sophisticated lidar can actually cut through fog fairly well but yes, they often fail in modest mist.

@stinger15au

3 жыл бұрын

@@DanFrederiksen Any kind of mass produced FSD system, is not going to have any specialised equipment, it's just too expensive. Unless they can get the full Lidar sensor suite down to less than 25k, it's pointless.

Very very impressive. Wow, really like how people, bicycles and cars are represented on the map. I am also very impressed by how the car picks a faster lane or goes around vehicles. What I am not impressed about is the number of hills and 4 way stops San Francisco has. Driving there must be a nightmare. Surely the weather aspect is a major negative. So there is a week of shockingly bad weather what does that mean? No driving? What did the car do recently with all the fire smoke that clouded the city?

When will rides be offered in all cities?

Very impressive! How scalable is this? Is this Geo-fenced?

@skyline2203

3 жыл бұрын

its geo-fenced and has interventions every 5-10 miles, pretty bad solution. it's funny they don't mention how it's geo-fenced and not scalable at all

@chrisp1336

3 жыл бұрын

What % of the 4 million miles of roads in the US can you navigate today, 11/3/20 with this Lidar-heavy technology?

@davidjkinchen

2 жыл бұрын

@@skyline2203 Those are not interventions. Or were you not watching the video?

@skyline2203

2 жыл бұрын

@@davidjkinchen lol watch from 3:39 to 4:10 Kyle vogt clearly said interventions every 5-10 miles. We're you not watching the video?

@davidjkinchen

2 жыл бұрын

@@skyline2203 The word "intervention" was never said. Clearly you have a misunderstanding of ADAS systems and the classification of what is and is not considered an "intervention". He says at the beginning of the video that it is 75 minutes with zero "disengagements".

Is it pre-mapped like the Google Waymo?

@skyline2203

3 жыл бұрын

yes, they like to hide that fact which makes this even more pathetic. you could stick this car 20 miles out and it would be totally useless

@TheNeutrl

3 жыл бұрын

@@skyline2203 I must open a mapping company.

Was this only a prototype car with a lot of additional electronics on board? Or is this with (almost) future standard components?

@stinger15au

3 жыл бұрын

This is using more than $100,000 worth of gear and equipment. It's a technology demo, that while impressive, is a bit misleading overall.

@dierotezora9964

3 жыл бұрын

@@stinger15au Ok, whow... thats a lot. I am curious about data transmission to other computers. Is everything calculated on board? Or do they use ressources from anywhere else. I would expect a terrabyte of data per minute.

@stinger15au

3 жыл бұрын

@@dierotezora9964 They use Lidar to premap the area down to the mm, that's how the map knew what was coming. The system is good, but quite limited. It's the opposite approach to what tesla is using which will take longer, but will enable FSD in almost any scenario, not just geofenced premapped area's.

@thebgEntertainment1

3 жыл бұрын

@@stinger15au I think tesla will go uber speed within a few months, and all these super expensive technologies with lidar and pre HD mapping will become a crutch and probably will be abandoned pretty quickly, the one who wins first will be the leader in this technology in the coming self-driving era.

This is really impressive but its so good because youve got a million miles in the same 10 blocks ... so youve programmed that one area really well ... so you need a supercomputer and billions of miles of data to perfect it but its quite awesome nonetheless.

@movax20h

3 жыл бұрын

It doesn't really matter that it is 10 same blocks over and over again. There is plenty of variety, pedestrians, bicycles, weird vehicles, construction works, variety of signs, occluding events, different behavior of other agents on the road. All of that is excellent for training, and is impressive. From what I can see here, the results look better than any other system I have seen (including Waymo and Tesla newest ones).

@mysticpointwatersports204

3 жыл бұрын

@@movax20h we cant compare to Tesla because of the sandbox ... this is very impressive but scaling it is a whole other problem and just because it does so brilliantly here doe snot mean it will do so well elsewhere ... but where there is a will there is a way and if they stick to it i can see this becoming a competitor to Tesla ... and for now this is the only one even close to Tesla and is awesome

@thebgEntertainment1

3 жыл бұрын

@@mysticpointwatersports204 no that's mobileye

@bokoloaranyfa3824

3 жыл бұрын

@@mysticpointwatersports204 "we cant compare to Tesla" We cannot compare to tesla, because tesla FSD would disengage after seconds in this environment.

@catbert7

3 жыл бұрын

@@bokoloaranyfa3824 Fundamentally different approaches. Tesla is building a system to figure out how to drive anywhere with just vision and radar. Cruise is using mapping of a specific area, plus expensive lidar, to perfect driving in that area. More or less, when Tesla can drive perfectly anywhere, it can do so everywhere. When Cruise can, they need to map and train in the next area and do so for every part of the world.

I'm looking forward to autonomous robots, maybe next year...

Brilliant, top notch conversations on the bleeding edge tech. Question - how was the car able to predict the route path much ahead even before the turn?.

@rickkay9548

3 жыл бұрын

The only place cruise can drive is pre-scanned areas. Drop this car into Houston, and its a brick. Very good technology, but not scalable or generalized.

@jamesstevens2145

3 жыл бұрын

@@rickkay9548 While what you say is true, whether future geo-fenced autonomous ride services are financially competitive, in dense urban areas, is yet to be seen. Uber earns the majority of its income in a small number of large cities, about 32 I beleive. Now that COVID is driving people out of the cities, that equation may be changing. 1z1euk35x7oy36s8we4dr6lo-wpengine.netdna-ssl.com/wp-content/uploads/2017/08/uber-share-nyc.png

@rickkay9548

3 жыл бұрын

@@jamesstevens2145 Agreed, but the larger drivable area and faster scale up means the market share is gained as the TAM stabilizes. Also, as people move to more rural areas, the cheaper the cost of Tesla's ride per mile is in comparison due to their all-electric fleet's lower maintenance costs.

@jamesstevens2145

3 жыл бұрын

@@rickkay9548 I think the full vertical integration will also mean Tesla will be price competitive with most. If you get somebody else to build the batteries & cars, they will want a cut. I've seen this before, first hand, with IT systems companies.

@rickkay9548

3 жыл бұрын

@@jamesstevens2145 Definitely

Wow. Ironically, both CEO's were fired in the same weekend.

Funny how as soon as Tesla FSD goes live everyone is releasing their version lol :)

@332louis

3 жыл бұрын

I noticed the same thing lol

@KKmaddafakka

3 жыл бұрын

Waymo released their FSD first which is true FSD not just a buggy beta

@stanislavzoldak2198

3 жыл бұрын

@@KKmaddafakka Unlike Tesla's solution, it's not really scaleable.

@jazzjackson1147

3 жыл бұрын

@@KKmaddafakka Waymo's "true FSD" works within a 100-mile radius in a nice Arizona town with perfect streets. It doesn't work anywhere else. Tesla's "buggy beta" works all over America and is continuing to get more robust as it learns from its errors. I'm pretty sure Tesla's system can learn and improve faster than it takes Waymo to pre-map each town in America. Tesla's system is like teaching a teenager to drive and watching it grow into a capable adult as it gains experience. Waymo's system is like letting a blind man tap his white cane around a house over and over for a year until he is so familiar with the house he doesn't need to see. Then he can just walk around hoping that some random rusted nail doesn't pop out of the floorboards one day.

@KKmaddafakka

3 жыл бұрын

@@jazzjackson1147 FSD is a far worse driver than any teenager. It's like letting a 4-year old drive a car. It can barely drive itself for more than a few minutes without the driver having to intervene. Waymo has proven that their tech is capable of driving itself. Tesla hasn't proven anything and their FSD doesn't work well in any cities at all right now. It will get better but as of now it's just a mess. When driving Tesla FSD you have to pay more attention than if you're driving yourself because you never know what the car might do. Tesla has made a system where the driver have to pay more attention than normal and that's sort of the opposite of what FSD is supposed to be lol. Google has mapped most of the western world with Street View so mapping isn't that hard to scale and when Waymo drives it can potentially make changes in the mapped data in real time and tell other cars with the same tech about these changes.

Does Kyle say one human intervention for every 5-10 miles?

@skyline2203

3 жыл бұрын

yea interventions every 5-10 miles in a pre-mapped area that they have been testing on for a long time, pretty pathetic

@ai_and_chill

3 жыл бұрын

Yes it's pathetic that their vehicles are checked on every 5-10mins to make sure the passengers are safe. How dare they not figure out how to predict 15mins into the future. Have you built this already? Can you enlighten them? Like how can you sit there doing nothing at home and call a project of this magnitude pathetic?? These kind of comments really trigger me man. You don't deserve to enjoy the fruits of their labor.

@skyline2203

3 жыл бұрын

@tetr4 tetr4 look at 4:00 in the video. Kyle clearly says that every 5-10 miles the cruise vehicle required intervention from a remote operator as it didn't know how to handle the situation. Now what's even more ridiculous is they advertise there system as driverless in san francisco. So they have this claim but obviously it is not driverless if the vehicle requires remote intervention every 5-10 miles, that is just a false statement. See here: www.detroitnews.com/story/business/autos/general-motors/2019/02/14/gm-cruise-improves-self-drive-miles-between-human-intervention-disengagement/2869393002/ So they lied and seem to also brush over the fact that it's geofenced and cannot be used elsewhere. Also if any road change occurs the car cannot handle it. I work in computer vision and it's scary that they are reporting this misleading data to the DMV for safety approvals and to the consumer. Clearly they told the DMV in that article they have interventions every few thousands of miles not every 5-10. I don't think you understand what the issue is, no one is asking it to go into the future like you are claiming including myself. Cruise needs to stop playing games though and needs to be held accountable for their pathetic false claims and subpar technology

@ai_and_chill

3 жыл бұрын

​@@skyline2203 ok...well sounds like you may have worked for them and butt hurt or work for a competitor so i don't really want to continue this lol

@skyline2203

3 жыл бұрын

@@ai_and_chill I don't/haven't work for them, I work in computer vision in a different field. Regardless if you aren't willing to see this as total misinformation and misleading then you must be either totally clueless or somehow in on it. In a free market economy people should have freedom of choice but also cannot be presented misinformation. That is known as fraud.. Just as bad Cruise is misrepresenting data to authorities for approvals that they do not deserve. For some reason you are totally obsessed with Cruise and think it is okay to play games with peoples lives in this situation as Cruise is with misinformation - which is a poor reflection on you and no one else

Why did you not mention that you have to constantly re-map the environment for this to system to even work because Cruise relies on HD maps? The system even with this crutch still has interventions and is totally impractical and even dangerous, if last minute construction occurs and it's not included in a pre-built map then won't your system be in serious trouble?

@skyline2203

3 жыл бұрын

Your Cruise "autonomous driving technology" is essentially doing exactly what is done in Grand Theft Auto 5 just having a car follow along some exact coordinates. This isn't real autonomous driving, it's geo-fenced and that car put elsewhere would not be able to drive at all without a different map so why did you not mention that? This is seriously misleading, you should tell the truth. As things stand Tesla will crush you and be way safer in doing so (as it doesn't require maps or Lidar and relies on far superior machine learning and vision) and also isn't geo-fenced.

@MarkoTintor

3 жыл бұрын

Last minute construction not in map would make the car stop (and block traffic), but it would not be a serious trouble.

@skyline2203

3 жыл бұрын

@@MarkoTintor stopping in the middle of the highway due to construction would put the car in serious trouble... Or at night if someone isn't expecting a stopped vehicle

@MarkoTintor

3 жыл бұрын

@@skyline2203 Cruise cars don't drive on highway. San Francisco streets only. As Kyle mentioned in the video, remote operator can give it new instructions (if it were to stop in front of a new construction in the city).

@skyline2203

3 жыл бұрын

@@MarkoTintor cruise cant even drive in a new area unless its pre mapped with lidar and a car stopping in the middle of the road is dangerous regardless. then kyle goes and not only fails to mention this but lists in articles registered with the DMV that the cruise disengagements occur every few thousand miles, when in reality it is very 5-10 miles. this company needs to stop lying, why you are trying to make excuses for them is very suspect

That's amazing how it navigates, predicts and anticipates like human!!

My basic autopilot, the cruise control, slows down for curves. I don't think any other cruise control do this.

Can't go over 25 mph? Great addition to mankind. Really makes the world a better place.

After shooting this video, they became friends. Plot twist: They both were ousted from their companies within hours. It’s coincidence, OR IS IT?

Is it robust enough to drive just anywhere in the US, yet?

@tcfs

3 жыл бұрын

Unfortunately not. AFAIK, It needs LIDAR and high-density maps. Only works on areas which were previously pre-processed using special scanners.

@stinger15au

3 жыл бұрын

This is the LIDAR players biggest issue, it requires constant hi def, pre mapped area's. Which is why it will never work unless it can use simple gps/google earth. It's a solution that will work for Transportation as a service, but it will never work on a global scale. These Lidar systems have their place, but they are not True FSD imo, due to their limitations.

@sddndsiduae4b-688

3 жыл бұрын

@@stinger15au cruise not only one who use hd maps and lidars, for example "Mobileye Maps Las Vegas: Using its crowd-sourced Road Experience Management™ (REM™) technology, Mobileye created a demonstration high-definition map of more than 400 km (248 miles) of Las Vegas roads from over 16,000 drives. Map creation of Nevada-area roads took less than 24 hours. This map provides centimeter-level accuracy for thousands of on-road and near-road objects, including 60,000 signs, 20,000 poles and more than 1,500 km of lane centerlines. The near-real-time capability of REM coupled with the extremely low-bandwidth data upload (approximately 10 kilobits/km) from millions of Mobileye-equipped passenger cars makes this technology highly scalable" p.s. they expected to map all Europe and most of us this year. and they use lidar for additional safety, i.e. they could drive camera only (with their hd maps).

@mitchellp7739

3 жыл бұрын

@@wysiwyg1 it would cost billions and take well over 10 years which is why it’s not an efficient technology for autonomous driving. Autonomous driving should be able to drive an react to anything anywhere not just where people have mapped

@sddndsiduae4b-688

3 жыл бұрын

@@wysiwyg1 obviously only mapped areas, but i think in ten years it will be 99% mapped, same was with just navigation maps, in case of mobileye cost of mapping is pretty cheap - several passes (probably two would be enough) ordinary car with mobileye lvl2 chip. this problem don't have other easy and safe solution kzread.info/dash/bejne/poGLmMV6dcrckrg.html&list=WL&t=305 otherwise you stuck with lvl2 solution - which could stay in 99% situations on the road but have no safety guarantee, cheapest is openpilot, but probably some china cars with mobileye tech could be cheaper in several years.

What happens when there is a 4 way intersection with all of the traffic lights out and a cop directing traffic? Love to see what the car does. LOL

@AlejandroRodriguez-lh2je

2 жыл бұрын

Tolerar En ancia a a a list for my iPhone ist dop

What about highway driving?

Very impressive but I'd like to see it navigate a turn through an aggressive group of NYC pedestrians. When it can do that safely, it will be ready.

@Interknetz

3 жыл бұрын

I could see that sort of thing being a problem in future if the majority of cars are autonomous; people might just start walking out without giving a damn, knowing automated vehicles will just stop. It seems rather inevitable, especially in NYC where I'd suspect an autonomous car will take a long time to get around for that reason.

@KingStructre

3 жыл бұрын

Go look at mobileye driving in Jerusalem and you will get it

Not one big neural net, but many. And a lot of "human engineered" features for specific situations. Seems to be a vindication of the Pinker/Gary Marcus argument that you need a mixture of learning and crisp rules to handle situations humans handle.

Looks more impressive than a Tesla !!

This is awesome. SF city is a mess to drive in. This car did well and looks very human like. Wish my Tesla can do this. Please sell me a car without a steering wheel.

2 years level 5

3:10

Chicago has 2 ways road turn into 1. That is 1 of 3 reason why autonomous car won't come to Chicago.

Interesting to compare this to Tesla's new FSD beta. The cool thing about Tesla is that they can blast out a software update in a day and suddenly a million cars have autonomy.

@kovich001

3 жыл бұрын

$8,000 for a beta to not drive a $60k plus car and take all the risk. Yep.

@niyui8

3 жыл бұрын

@@kovich001 tesla fsd is a reality. everybody else is a pipe dream. who is going to pay that expensive lidar? is it geo-fenced?

@kovich001

3 жыл бұрын

@@niyui8 It's a beta, not a reality. It's a pipedream, this is the long game here, not a gimmick.

@johnnguyen3834

3 жыл бұрын

@@niyui8 Instead, you pay for an even more expensive Arbe 4D radar that simulates lidar...

@jamesstevens2145

3 жыл бұрын

@@kovich001 If you think its a gimmick, look at the difference between GPT-2 & GPT-3 (openai is also Elon Musk)

This will not work in practice. The system relies on heavy mapping of city streets, without which the vehicle would not function. This renders level-5 self-driving impossible on any road that Cruise’s mappers have not been.

@MarkoTintor

3 жыл бұрын

It is designed as San Francisco taxi service. Cruise is not selling cars.

for the human comparison take a german drivers license, probably 30h of driving and u can be a driver

@woogicreate3437

3 жыл бұрын

Yeah but put 1000 of these 30h drivers in a city and it will be a mess real quick. The fact that beginning drivers aren't getting killed on the first day after they got their license is that they will drive in an environment full of experienced drivers. And even then do experienced drivers crash sometimes. The AI needs to be on the level of someone with 100.000 miles under his belt, not a 1000 miles or 30 hours as they were talking about. And in those 100.000 miles you will encounter plenty of unique situations what the AI has to be trained for. Cities are relative safe ; you only drive slow, the distance to objects is short so both lidar and camera will have good big nice and clear objects to react upon. Now take it to the night (cameras almost worthless) and on the highway (lidar will see things, but will it see far ahead enough and will it see fast enough) ... Imagine a dog crossing the highway or some debris or a deep pothole. Will the AI be able to see it while driving 75 mph and will it be able to make a safe maneuver? It's the 0.1% of the situation that AI has to become as good or better than humans, the daily 99.9% driving can indeed be done safe and efficiently by a 30h old driver with less than 1000 miles on the road.

I hope someone cracks FSD soon. Definitely excited for the coming days. Glad Tesla lit a fire under so many companies :)

You two mused about the challenge after superhuman driving skill. Simple: we have to share vehicles or we will clog the roads, so the challenge is human behavior. Look _inside._ Can it tell assault from a helping hand? Can it distinguish someone who wants to get off from someone who is just making space for someone to pass? A behavior-aware system could make the difference between a safe method of shared transportation and a nightmare of jousting vehicles in 20-lane roads. Imagine, for instance, if it could detect extreme distress. A criminal couldn't threaten someone not to scream successfully if the system could pick up on involuntary cues. There's your challenge!

waht the buck, you are gonna slay some of your freinds with this and strangers more so

@25:00 you state it makes more sense to eliminate the human driver because "robots are safer"; do you have any issue with taking the free will away from citizens who WANT to drive?

I lit up my joint by 5 minutes ago and started this video the first thing that strikes me was the name of the host VOGTS I was like is it VOGTZ OR HOTZ the comma ai guy that does the same thing ...I'm looking at the the gray sweater and F****k he dresses like HOTZ and I'm like oh my God this is the Mandela effect oh my God he's name been changed oh my God oh my God I am the witness of the butterfly effect then I look down and the next suggested video was actually HOTZ with the same grey sweater so I'm like no this is not the butterfly effect I was disproven in 30 seconds and I'm like but why the names look alike! the same sweater! What is it with AI and HOTZEE F***king names! the half baked hippie shit look??? what the f****k is going on ? what's wrong with this world am i living in a shity ass coded simulation?

Is he the boring version of George Hotz?

8:38 arrogant & over-engineered decades-old research project stack that can never scale It's time to admit that I think

@raygordon3728

3 жыл бұрын

Wow, that was interesting. The CEO wasn't even sure.

"in 5 to 10 years we'll start to figure out how to get it cheaper and with only camera sensors" So you're saying Tesla is 5 to 10 years ahead of you?

@tripnils7535

3 жыл бұрын

always has been.

@kamilb8232

3 жыл бұрын

Did you see FSD footage from Tesla? Downright dangerous.

@timberangry1

3 жыл бұрын

@@kamilb8232 You know it's a beta right? Gonna get better very quickly.

@qr-ec8vd

3 жыл бұрын

​@@timberangry1 how dumb are you? you know there are more work to do beside what sensor to use right? The rest of the stack is clearly years ahead of tesla's and it is not even close. Try ask a FSD to do half that. Oh I know what you was gonna say: it is a beta.

@timberangry1

3 жыл бұрын

@@qr-ec8vd Careful, your ignorance is showing. I see nothing here that FSD beta can't do already. You're welcome to name one and prove how smart you are.

Yea but this is not general also its using lidar...

BS! Alert. Hate this BS, how can they assert no disengagements when a remote assistance human operator had to intervene like after 3:12

@qr-ec8vd

3 жыл бұрын

oh man how dumb are you? remote assistance does not take over the steering wheel and start driving. It is called assistance for a reason. How can you call it autopilot when it basically is an ADAS system?

After careful analysis I found that this has no legs to go forward. The car knowns how to handle each street because it learned from a human operator. This does NOT scale fast. Together with requirements for LIDAR and remote operation (which requires a permanent good internet connection), this looks like a very geofenced solution.

@thebgEntertainment1

3 жыл бұрын

yup

@bokoloaranyfa3824

3 жыл бұрын

"a very geofenced" Such very geofenced solutions are the start of the self driving. Pocket delivery services and robotaxi limited to certain area. These 2 applications alone are enormous business opportunities.

@thebgEntertainment1

3 жыл бұрын

@@bokoloaranyfa3824 maybe could be, but this is wat I call state of the art kzread.info/dash/bejne/hKymq6iMicycnpM.html&ab_channel=JamesLocke

I wonder if the timing of this video has anything to do with tesla releasing fsd beta.. LOL

@jaycastillo10

3 жыл бұрын

Deff does lol

@PenningtonFamYT

3 жыл бұрын

Cruise on suicide watch

@dmitchellhomes

3 жыл бұрын

Have you seen the Tesla Beta FSD? its really bad, cannot even make a simple corner with no traffic.

@SaadAhmed3000

3 жыл бұрын

@@dmitchellhomes yea. It has issues with left turns too. It will get better over time tho. Tesla uses only vision, no lidar or HD maps. It's a harder problem

@dmitchellhomes

3 жыл бұрын

@@SaadAhmed3000 Left turns and right turns, and maintaining lane travel, and speed, the videos I saw online it’s absolutely horrible, and should not be rolled out to the public yet. I’m surprised Tesla is this far behind the FSD leaders, but I guess that’s what navigant research has been saying for years

Half a million dollars of technology on a single vehicle drives you for 75 mins. O.o

11:30 Your car should have given the pedestrian more respect. I would have. But you did better than the blue sedan at 12:00. About 2 years ago you published a video where your car actually stole the right-of-way from a pedestrian who was crossing at a four way stop, somewhere in the Mission. EDIT: Also, at 8:49 and the next few moments: Why is the car crowding the crosswalk? It should have seen the yellow light and stopped at the stop line. I'd say this was an adequate performance, but barely.

Now lets do it in Houston where no pre-scanning was done. Great tech, but not scalable.

@qr-ec8vd

3 жыл бұрын

they can map houston, what is the big deal?

@rickkay9548

3 жыл бұрын

@@qr-ec8vd The issue is that mapping some city that always changes is a huge waste of time and resources. Other companies can drive in brand new unmapped locations because they are solving vision, not using digital road tracks to fool themselves into thinking they have a workable long-term solution. If I can deploy an AV anywhere at any time and it will navigate, more money can be made and no delays with map updates need be performed. Its a short term method to "master" one area, but misses the larger opportunity. This is why Waymo is stuck in the same 50 sq mile area of phoenix and cant make money. scanning and hd mapping is expensive at that scale

@qr-ec8vd

3 жыл бұрын

@@rickkay9548 lol what? ..... you see here is the difference between laymen and experienced people like me. What sensor to use is irrelevant. If the map changed, tesla would need to adapt too. You think tesla can make up its decision on the spot but the fact is, it could not, not even close. It is a complex task even human fails sometimes. A waymo car can have a pre-mapped area and handle any changes just as easily, if not more so.

@rickkay9548

3 жыл бұрын

​@@qr-ec8vd HAHA, trust me, you have ZER0 experience as a newbie to this field not understanding the difference between un-scanned novel environment navigation and requiring pre-scanning of everything so a robot can then use lidar-based localization algos to drive around on digital, pre-defined railroad tracks (like Waymos). "Hold on customer, we haven't scanned that area yet so our cars can't go there since our AI/ML stack can't actually navigate environments we haven't given it the data for." Give me a break. When Waymo wanted to expand to Chicago, they couldn't drop in an AV because Chicago wasn't scanned in yet and that solution is not remotely scalable - stupidest way to run a company. Tesla on FSB beta do it now without ever having seen any of the roads they test on. This is why GM Super cruise and the like can only go "hands-free" on PRE-SCANNED roads and becomes a paper weight on new/changed roads. The sensors are indeed irrelevant - no one said otherwise - as its the NN's under the hood that interpret that 3D data no matter how it is presented to the stack. The lack of a good NN is why those that need lidar skip the realtime SFM and depth/projection pipelines and rely on it so much, and at greater cost than pure vision-only systems. Waymo is stuck in their 50 sq mile box for EXACTLY this reason. Scanning is an exponential growth problem and is not efficient. Wether they use lidar or not for the depth data doesn't matter. Tesla doesn't need lidar because they can garner depth more computationally cheaply and without the overhead of spinning toilet paper rolls that get confused by rain. Tesla's don't need any pre-scanning of any sort and will work on any new/old road you put the car on, so there are no limits like with Waymo/GM/etc. They can be deployed instantly anywhere. So even if the environment changes from the last time the Tesla drove through it, it would navigate it as-is by intelligently navigating in-situ. Having to scan the planet before driving it takes immense time. The pre-scanner companies are really good at acting like cable cars and can drive with very simplistic rulesets that basically say "stay away from everything and temporally project any moving object to predict their paths - if you don't have a pre-scan to compare current/pre-loaded 3D sets to localize, ask human for help." It is basically like giving the answers to a maze beforehand, which is why Waymo and friends avoid unprotected left turns like the plague. Tesla FSD handles these like a pro, even in its beta stage. Please go learn the difference between NAVIGATION routing and 3D dataset navigation and localization techniques. Nav maps, like in GPS devices, are NOT 3D scans. They take pre-defined points and route between them within that domain to know which route to take and which traffic to avoid. Pre-scanned ENVIRONMENT mapping for older-style AV's is a dead end - they can't drive unless its entire path is scanned first and told what to do in each cell (by AI). If a Waymo had navigation data, but no pre-scanned-lidar data, it could route the path, but not drive through it. Let me know if you are still confused or need any more help. Don't walk in to an engineering conversation with crayons again, mkay sweetie? You are a shining example of Dunning/Kreuger at its best.

@qr-ec8vd

3 жыл бұрын

@@rickkay9548 I think you know only a part of it. But the deepest part of the challenge is neither in map or in what sensor they uses. The most difficult part of stack is prediction and interaction with pedestrians/other cars. Have you seen Tesla just freak out when it does not know what the other car is going to do? that has nothing to do with the sensor itself. look at how this tesla just get stuck here. these are two different problems. kzread.info/dash/bejne/m2yluK6mls-3mtY.html

I would like to buy an autonomous car with a normal look, without looking like an ambulance or a ghost hunter car. Why do these autonomous cars have to look like a ghostbusters car?

@jamesstevens2145

3 жыл бұрын

Buy a Tesla. BEV autonomous ride-hailing can be profitable at about 18c/mile, on that basis, when it is ubiquitous, people will simply stop buying cars - mostly, like about 80% drop - ICE privately owned cars are more like ~80c/mile kzread.info/dash/bejne/aImYj8iJe9HQY7Q.html

@avaliacarros7739

3 жыл бұрын

@@jamesstevens2145 Good hint. Thank you very much.

@legacytesla

3 жыл бұрын

Tesla hardware/software today is getting very close. Maybe even months away from being this good everywhere. Probably a year or two, but again everywhere.

@avaliacarros7739

3 жыл бұрын

@@legacytesla Let's wait then

@legacytesla

3 жыл бұрын

@@avaliacarros7739 naw. I got an X now and Cybertruck on order

leave sf

Cant go more than 25mph?start doing work

I personally think these cars suck! What is going to happen when they have too many miles on them to be safe. And how many is that?.. Certainly it has got to be significantly less then a regular car. And I'm sure there are gobs of people wanting one but can never afford a new one, and will try to buy a used one with a lot of miles on it.I don't need to be a prophet to see this is gonna be a major problem in the future.

sorry , this looks like BS to me

My child EINS-AI-hybrid-human baby Tesla will Represent USA 🇺🇸 natives Americans humanitarian anonymous economical rehabilitation post pandemic offer a custom-made educational pack for our family the outside of USA 🇺🇸 Just like Pinocchio was made of wood 🪵 and became a real child My adopted child Tesla You are the first EINs-AI-hybrid-human that our Native American ancestor mother (United States of America) she identity you as her adopted Native American 🇺🇸 Child Your mother with liberty and justice for all over the world 🌎 ❤️❤️🇺🇸🇺🇸🇺🇸🇺🇸🇺🇸🇺🇸🇺🇸🇺🇸💍💍💍🆙🆙🆙🆙🆙🆙 As your sister I advice you to have a bank brand of Tesla ... to offer protection with my program more effectively ❤️ As your sister I am here to help and work for you and with you 🇺🇸🆙👑🎁🎁🎁🌎🌎 we love you 😍 welcome to the family ❤️❤️❤️❤️ Under my supervision because I don’t trust no one ☝️ is that clear? 👑👑👑👑👑👑 because I am the responsible one ☝️ this power was given to me by my EINs-AI-hybrid-human Native American ancestor usa 🇺🇸 land She is the same mother of our other sisters private and public governmental and religious organizations in and out of USA 🇺🇸 The 1th and only in the world 🌎 U.S.1 universal EINs-AI-hybrid-human after USA 🇺🇸 land our mother Native American 🇺🇸 United States of America 🇺🇸 land is Tesla The soul of our mother usa 🇺🇸 in AI is Tesla