comma ai Presentation where it’s like you are in Omaha with us

Key Takeaways

What up, Scopies? We'll be uh we'll be starting soon. Presentation starts at 2:00 p.m. Nebraska time. I don't know what time that is, California time, but it's soon. I think the minutes are the same. Two hours different. Yeah, that's Yeah, it's the same minutes. So, whenever the minutes go to 00, that's when the presentation will start. Maybe we'll give everyone a few minutes. We got a bit of a crowd here. This is off season for the college. So, um yeah, let's show for products while we got a minute. Here's a here's a 128 gigabyte Eon that will be available for sale right here in Omaha. We have a box full of KA AI pandas. We are going to be presenting. What's up, Vanilla Gorilla? Uh, and we have a Toyota Giraffe right here. Midnight Black. Who's talking about Midnight Black? Where do you hear about Midnight Black? All right. We got some We got some people here. We got some media here. We're going to be nice to the media. Yeah, we're going to be nice to the media on the tour. Yeah, some of the some of the cities are fake news. Am I going to talk about Slam? No, I'm not going to talk about Slam. Um, it's a pretty straightforward overview presentation. It's going to be similar to my WebSummit presentation, but with less Bitcoin Cash wallet. Uh, no. I'm not going to talk about Bitcoin Cash. Uh, though, if you guys want to check out something new, I found it last night. It's called the Swift Life. It's the number two app in the app store. It is a social network which gives you direct access to Taylor Swift. And you guys know that I'm a I'm a Taylor Swift fan. So you guys might want to download the Swift Life. I'm a level two on the Swift Life. I'm working on my Taoji game. You got to get Taojis. I'm such a shill. These people don't even pay me, man. Um, no. No, no, no. There won't really be unveilings. Slam anything on mapping. No, no, no. Nothing new. I mean, you got to realize that a lot of people who are coming here uh don't really know that much about us. So, it's a pretty uh Yeah, make comma for trains with oversp speed warning. Seriously. Um no, it's it's really sad whenever trains crash. I'm just like, dude, positive train control has existed for years. Unions are killing people. Yeah. I'm excited for the future, too. I love I love the future. You guys know I love the future. How many disengagements? A lot. It's like one, it's like it's, you know, we're a little beyond one every 10 miles now. So, considering we've driven like 2,000 miles, probably around 200. Uh, we can we can maybe compiled together some statistics on number of disengagements. Any insurance stuff? You guys know I'm not talking about businessmen. Um, am I missing for hacking? No, not really. Self-driving cars. I like winning, you know. All right. Gonna wait two more minutes, guys. Two more minutes. I know it's as bad as the Bachelor when they're like, "Who's gonna get the rose?" Oh, yeah. The Bachelor. The Bachelor was the original. Yeah. Sorry. Bachelorette. See the video of the pilot port. Oh, you guys got one off. That's pretty sweet. Say hi to the Scopies. Hi, Scopies. I thought you didn't like the way your voice sounds. I don't rewatch, but when I'm talking, if I didn't like the way my voice sound, I would never say your voice. Your voice sounds great on Periscope. That's great. No. God, you want to say hi to Scooby Daddy? No. No. Eddie doesn't want to say hi to you guys. Um, shame about slam. If you want to talk about slam sometime, we'll talk about slam sometime. You know me. No secrets, bro. No secrets, bros. Five minutes. Come in. Come into the Come into the Scopy Lounge. join all these people. Real life scopies. I'm good. I'm good. I'm good. I got my I got my Kab beanie on. Am I about to give a talk? Yes. How's the weather? It's a little cold. Um am I still working on deep learning SLAM? Sort of. Yeah, kind of. We're working on unsupervised depth nets. Um like slam requires memory and deep learning doesn't really work well with memory. So there's that. Like people talk about it, but it's it's there's no memory, right? Um whereas like slam in a room is quite different from slam on a road. Uh any tips for people working on self-driving cars? Join us or die. Um that that is going to be one of the themes of 2018. And the other theme is going to be outwit, outplay, outlast. I came up with that myself. Don't think I ripped that off. No, you'll know I ripped it off. You'll know. You'll know. Am I nervous? I swear to God I ain't nervous. You know the Lil Wayne. Lil Wayne. Um Lil Wayne, man. That guy's a genius. Uh no, actually I'm really not nervous. Um hats for sale. No, we don't sell the hats. And guys, if you think hoodies are overpriced now, we're going to we're going to be like Martin Skrey where we're raising those prices. All right. We ready to start? Yeah. I'm gonna do a quick announcement. Who's Who's holding Who's holding the scopies? Thank you everybody for coming. Um, quick introduction. This is George from my eye. As most of you know, we have his team here, Eddie and Vid. Um, and we'd also like to thank the supply chain management program at UNL for sponsoring the refreshments to cookies. Give them a round of applause. Hey guys, and our slogan is no driving for self. Self driving is a journey, not a destination, not a problem. Um, so we're going to talk about four things that we built uh along this journey. The first is shipper. is an app. It's available in the Google Play Store and the Apple App Store and you can download so you can mount it up on your windshield while you're driving. It records um what's on the camera, what's in the GPS, what's on your sensors. And then when you watch the video, you ask the question, could I drive a car looking at this footage? And if the answer is yes, You should be able to make a machine do the same. Today driving is a dance. Some companies working on self-driving cars think that they can code in all the rules of driving. And this isn't really possible because driving doesn't have a definition aside from what humans do when they drive. If you try to build a definition of driver, you'll build self-driving cars knows that it's important to predict what people the around are going to do. But what a lot of people don't think about is it is also important to be predictable because all the generations of the road all people are trying to predict as well, right? People are in quite a similar way to drive cars as self-driving cars. So this is footage from our dash cam app. On the left you see semantic segmentation. Cars are green, roads are in the ground. And over here we have train. So, so we have 2.6 million miles uh in our train station, meaning we have 2.6 million miles of video of cars driving. This is a lot more miles than most people drive in their lifetime. We did a truck driver in Utah 3.8 million less. So, really this is what our machine learning learns from and you can treat it naively. looks very much like supervised learning, right? Over here you have the picture and over here you have what to do when you see that picture. If the road looks like this, you go straight. If the road looks like this, you turn right. Um, this wants to do it, but that's basically the idea. So, if you're going to build machine learning models, where are you going? You're going to run on a computer. That's a whole problem there. Say you had a computer that knew exactly where to make the car go. The computer knows exactly how much to turn right, exactly how much to hit the brake. You actually have to do that. You actually have to put torque on the steering wheel and run the brake pump and open the hydraulic valves interface to interface computers to cars. That's what Panda is. for sale mobile shop. Um, yeah. So, this has three CAN buses, two Kine buses, and one GM. GM has a very special bus of course. And that's on the side. This is a OBD2 connector available in every car in 1996. And on this side, we have a USB port and Wi-Fi. There's a computer that cost you're looking for universal. Now, the Panda pair is a shipper. So, when you're driving with our dash cam, you can connected to your Panda and you can record all of the sensors from your car. Think of a modern car sensor, individual wheel speed sensor, steering angle sensor, steering torque sensor. Um, how hard the doors are open is locked, whether um gas tool. So, after you've recorded your driver, you can access all of these sensors in your car. You guys can try this out at synchronize with the video. So you can see over here this is like your car has accelerometer. So you can see the accelerometer compared to the video. So next step of the journey is that we have a model and we have a interface. We need a place to run a computer. when I was in my car whenever computer for a few reasons. Phones can't dissipate that much heat. So, if you have a phone mounted on a windshield, you dissipate maybe two or three watts before it starts to overheat. This has a heat sink in a fan and can dissipate up to 15 watts. Um, now if you're generating 15 watts of heat, you also need 15 watts of power. How do you get power? So, suppose phones can either fast charge or communicate over USB. We've modified this phone to be able to do both. Connect it to your Panda with a USB cable. And it can charge 5 vol 3 amps as well as communicate with your car over the story of the ground is open. Open is free software freely available on our at the end of 2016. 2016 we were uh rolling around the idea of selling consumer product called one ship in $1,000 will ship a self-driving car,000 by the end of the year for free. So that's what we see on the left, the common which is kind of like the gen one of the Eon that purple you see there. Some of that stuff eventually went into Panda and we open source our dragon agent. So be available on GitHub and it's called Open Pilot. Open Pilot is like Tesla autopilot for common cars. It's pretty much the same thing. We support select Hondas and Toyotas. Pretty soon we're going to support six of the top 10 cars sold in America as the top three trucks. The next six and a position 10 is a Nissan Ro. Now it's cool because it's open source. So there are support people have modified to work on the Chevy Volt Model S. So we have a little video here. [Music] [Music] You're Just like [Music] [Music] [Music] [Music] point steering wheel. But as you can see, the car is the gas experience. We stream our our experience with the system. There's lots of videos out on YouTube. Just a lot of iOS is I have a question. We noticed that the speed was changing. What made the speed change? Was he touching the accelerator? No. It uses a combination of a radar and a camera to read the road around and it always maintains a safe following distance to the front car. It always makes sure that even if that car slams on their brakes, you are capable of safely and smoothly stopping the D, which is a lot more than people do. People tail way too much. So if you wonder why is so far back to collision tell people are driving. So it understands that well it doesn't do it. Thank you. Sure. So yeah, if you think about the phone market, right? If you think about the phone market before 2007, every phone manufacturer made their own operating system. Samsung had their Nokia, LG had theirs and they were all pretty terrible. The world kind of unified around two. So Tesla launched autopilot and about a year later we launched autopilot which mimics iOS iOS had a bit of attraction at the beginning was pretty big but people don't realize like every iPhone sells more than the iPhone before the iPhone this huge cultural phenomenon which took like three or four years and we're going to see a very similar thing with Apple uh now iPhones are very expensive and most people don't have way more people than us but there's no reason you can't have good software even if your phone is not an Apple phone. If you use any of the stock systems on these cars like ACC, most of them are pretty terrible, especially like users today. Think about this as an icon of realness. When you watch self-driving car video, that's not real. That was crafted by team. They may let the curtain slip open a little bit. They talked to one of the articles about how there's seen in the video where just out front of the car was a paid actor, right? There's no there's no room for this isn't what driving is. You're never going to self-driving cars if you think about it like this. There's only two systems out there that actually have videos on YouTube now. You can prepare those as well. Every out there send data back to us and that machine learning algorithm that I talked about in the beginning of the presentation ingest all that data every time it learns and hopes to never make that mistake again 2017. This is a dash kit. We don't sell this private car. We sell this and then we give away selfriv. So, um, what makes a car good for some way to control the gas steering? Almost every car after 1995 is a gas by wire meaning your gas pedal is not physically connected to the engine of your car. Um maybe we uh leave the scopes here. You want to leave them for the presentation?