ai.bythebay.io: Stefan Seltz-Axmacher, Kartik Tiwari, Starsky Robotics: ↑↑↓↓← → ← → Autonomous
hey guys really started for keeping you waiting our talk today is about hacking autonomous trucks and apparently we should have thought more about hacking our calendars and not thinking our talk was tomorrow so really really apologize the Lexington team have been great in getting us set up and this is completely on us so we're Starsky robotics and what we came out of stealth last Tuesday as a driverless truck startup that's already moved Freight on the highway I wasn't in the 2004 DARPA Grand Challenge I am NOT an alumni of Google X we didn't go to Stanford by all means we're not supposed to be here we're not supposed to be a driverless trucks start off we're not supposed to have done this and the reason that we've gotten this done so quickly out of nowhere the reason a lot of you guys have never heard of us before is because we've basically hacked our way to doing this we didn't George Hotz hack a hack a dataport to get this done we took this incredibly hard problem that really smart people have described as being a ten year hundred million dollar type of issue to do and we've broken it down into something where almost everything we've built we've built since since I'd say last last June yeah yeah one of the other things that I want to point out is I came from another startup before before joining Starsky and well as you can see as you can see it didn't work out that well for me with my previous startup and one of the biggest takeaways from it was the value of Minimum Viable Product and we actually I actually carried that with me in into this one and you can see that by looking at exactly how we implemented the technology and and took a car check off to describe as last startup does anyone here like robots at all all right cool good crowd so part six last startup up droid we're building really really cool consumer robots so as a modular robot with a and you could pop on a set of wheels you could swap out sensor kits well weather and have camera rather have radar different actuators does it have one arm does it have two maybe you take off the wheels and put a series of legs all on a custom chipset that makes it really easy to do all these things and that's really neat but they sure as hell burnt through their 700,000 dollars of funding and didn't sell anything because while that would be great five ten years from now after multiple Kickstarter campaigns and multiple early iterations to launch that as your first product is really hard similarly the first start that I worked for was a privacy sensors social network where you could own your own content sell it to others using their digital currency and to top it all off they had a non-profit to protect your social network data maximum day they had 50 active users and 80 employees MVP is pretty key to how we've thought about this and we think that trucking is the minimum minimal Viable Product for the autonomous industry yeah and also not just trucking but the way we are approaching trucking is it's more or less the way we think is which is I mean it is the MVP of self-driving technology as such what just to go into a little more detail about how how we did this we looked at trucking industry and Steffan can talk about or talk more about the business aspect of it but as the problem came to me all we had to do was build a lane-keeping build lane-keeping vehicle and that can also do adaptive cruise control and we would we would already be in business so yeah so what's what's hard about self-driving being a self-driving car in San Francisco being cruise being uber for two days in December is really difficult there's a homeless person staggering into the street here there's a couple of hipsters walking across jaywalking over there it's a really complicated weird problem trucking is a different space so we didn't come up we didn't make a good algorithm and decide what's a good use case maybe trucking we thought about trucking first so many of you guys probably don't know this there is a severe truck driver shortage in America the reason for that is it's really really hard to get human beings to spend a month at a time in a truck if you're willing to do that if you're willing to live in a truck get home one day a month without a high school diploma you make sixty to ninety thousand dollars a year there's a shortage of fifty thousand drivers so the problem in trucking is not how great of a math problem can we solve the problem is can we make a truck move to make a truck move it doesn't need to be autonomous 100 percent of time to make a truck move it needs to move if you are building something that addresses this space without actually moving the person from the truck you're not solving the problem you don't have a product you're doing nothing which is why we're doing what we're doing remote control from a distribution center to a highway autonomous on the highway and their remote control to another distribution center long-haul driving is distribution center it's a distribution center at each end it's no more than five miles away from the highway we effectively made a driverless truck without ever needing to deal with how do we stop if a kid chases a bouncy red ball in Tamila Street and by definition very house districts in general are not not similar to neighborhoods in San Francisco so you already eliminate a lot of edge cases and because you already are building the whole technology on top of teleoperation as as its basic as its basic component you already have a sense of supervision that any any person in our command station can provide so that's that's the essence of the whole technology power platform the other thing is which sort of goes in into a political political scenario which is we also have human in the loop so we do think that like all the technology advances that that can happen in the next two to three years might make a fully autonomous level-5 vehicle possible but we as a start-up cannot wait for that to happen we don't have a hundred million we don't have 10 we don't we we don't have $100,000,000 we don't have 10 years we need to build something real when you build something now or we need to get jobs and we need a lot of data to build that level 5 car the level 5 vehicle to gather a lot of data we need vehicles on the road and we go into this vicious cycle of what our sorry chicken and egg problem for what comes first we so women we have already put vehicles on the road now and we about beer we're already gathering data while also doing business when Mo's spent fortune driving in circles in Mountain View over the last seven eight years and we also need a drive in circles since we're self-driving truck we should be driving in circles with some level of weight in the back of the vehicle to to actually simulate what it's like at different weights an unloaded tractor trailers about 20,000 pounds a fully loaded tractor trailers about 80,000 pounds so there's there's a lot of dynamics to learn there if we're gonna be driving in circles with weight we might as well get paid for it which all the Sun allows us to have the same testing footprint as as an auto as a much larger company mid next month we should have six to eight testing vehicles continuously gathering data for us which gives us one of the largest fleets and self-driving despite most of you guys never having heard of us before last week the other interesting thing about our approach is we are not using any technology or any product that has not existed or the shell that is not existed or in the market for the for the last five years so which is a nice way of saying we're not using lidar I've gotten cornered at parties about how how dare Tesla how their comedy I not use lidar everyone else uses it it's pretty telling that the one company really in market isn't using lidar it's a really cool sensor it's spinning lasers that's awesome I would love some spinning lasers the reality of it is that there are very few light ours that have been built that functionally operate at highway speed I've heard numbers that there's 500 to 600 of them they're hard to get ahold of if we were waiting on magic light ours to show up we'd be waiting to build self-driving trucks for a long time and four thousand people die each year in chalk accidents there's a moral imperative to bring safe technology to market and I'm not going to wait for the lidar companies to get their act together yeah I think that's a that's our unprepared doc for today and I'm sure we'd be happy to take as many questions as you guys have yeah let me grab the mic here for you guys oh sorry so you mentioned that you guys are going to have remote control from the warehouse to the highway merge stone what kind of communication protocol you can use for that secret sauce yeah this is a brutal industry we'll have some stories coming out later this year about how shitty human beings have been in this industry the last gonna half so that is secret sauce cool yeah and one more I'm very curious I noticed you guys took an approach of attaching your own actuators to the steering wheel to the pedals one why'd you take that approach and two how does that impact driver takeovers yeah so our system not on but with the whole system in the way as you might be getting at has been driven across the country multiple times so it doesn't have an impact so far that we've noticed on driver control of the vehicle the reason for actuators as opposed to going into the data port so when I first started working on this in September of 2015 I started emailing some like R&D labs for some giant OAM so I'm sure some of the richer ones of you own vehicles from and I said hey how hard would it be for us to just get access to your open API start issuing commands to your trucks la la Lele we have a self-driving truck and the response I got was well three to five years from us and I'd say civil amount of time from all of our tier two and Tier three suppliers so at that point the thought became either we wait three to five years which we don't have because we don't have three to five years of personal runway or we hack the truck or we actually just push pedals there's about a hundred years of engineering thought and experience on how to reliably push a pedal on how to reliably turn a wheel that and trucks unlike retrofit kits for Audi's don't need to look pretty they need to be reliable they need to be safe so that's why we are physically and manually actuating the truck as opposed to hacking into the data port and possibly compromising any of the safety critical systems it's also the single point of failure if something goes wrong we can actually back trace everything to either our our system or to either the truck there are like there are multiple answers to that question so if we're going in through the data port and the windows down and a splash of water lands just magically right which at scale will happen the data port fries out and you no longer have control over yourself driving truck on the highway that's pretty terrifying that's one point one of our one of our investors is a blake shoal from boom aerospace and we've been very much thinking about safety for autonomous vehicles in the same light and the same standards as safety for for aerospace so everything that we build has to have multiple redundant ways to safely operate the vehicle it's a lot easier to do that through mechanical actuators pushing pedals than it is through a single data port it's also a different different set of thought process when you look at the problem from deploying it as a business rather than being a science experiment big like safety does come first when you think of scalability and think of deploying these trucks on actual public road where other people are driving and what happens when nobody is in the truck and something bad happens so we that's that's why we actually take take like airline safety as our stand as I go standard and maybe like and hopefully surpass that that's that'll be our target cool probably have time for one more question if anybody has one oh we got one over here has to be on the opposite side of the room so with the remote-control aspect what actually happens if you lose communication with one of the vehicles whether it's some yep so as far as hacking there's been some clever stuff we're doing over there which is a nice nice cheap answer but as far as communication going down it is it is very far from black magic to make a self-driving vehicle pull itself over on the side of the road there's a decent amount of precedent for that so while working on this we've learned a lot of ways about how the trucking industry in general is incredibly messed up and and and and people are incentivized to drive them safely truck drivers can drive about 11 hours per day be on duty for 14 and when they're approaching that that limit they have to pull over and park somewhere but there's an interesting problem when you get to rest stops there's limited truck parking so frequently drivers will have to keep on driving past their legal and safe hours of service as they're falling asleep to find a parking spot which is why so often you see trucks pull it over on the side of the highway is it encouraged to pull over the side of the highway for an extended period of time absolutely not but it's regular enough practice that we can use it as a fallback position should communication get cut another way we're effectively able to to hack and and severely minimize the problem that we're going after is by only working with partners who need their trucks to go in certain places so similarly how do we deal in snow well right now we're not working with trucking companies who want to drive places where it snows how do we deal with communication and tunnels well right now we don't drive through tunnels that is not a good answer for when we have 3,000 trucks on the road but we're a ways away from that and that we're solving the problems to to get to market okay thank you guys