WEBVTT - The Race to Teach Robots How to Do Our Jobs 0:00:04.559 --> 0:00:06.680 All right, should I try to put something up so 0:00:06.720 --> 0:00:08.960 I would like so if you hit the space bar, 0:00:09.240 --> 0:00:10.920 you will be in control of the arm. You'll see 0:00:10.920 --> 0:00:14.200 your grippers now green. Um. You may need to move 0:00:14.240 --> 0:00:17.200 down a little farther because once you get to a 0:00:17.200 --> 0:00:21.800 certain height, it will refuse to follow you. Okay, um, 0:00:21.840 --> 0:00:24.759 and you would need your right hand on the mouse. Yes, 0:00:25.160 --> 0:00:27.639 that's me getting a lesson in a computerized version of 0:00:27.680 --> 0:00:29.800 a game I remember playing at truck stops when I 0:00:29.840 --> 0:00:32.120 was a kid. I'm sitting in front of a laptop, 0:00:32.280 --> 0:00:34.360 using a mouse in each hand to steer a claw 0:00:34.440 --> 0:00:37.600 around my screen and trying to grab plush objects. This 0:00:37.680 --> 0:00:39.440 is where your pointer is. Make sure you're not clicking 0:00:40.720 --> 0:00:44.080 buttons and so what you've just done is caused a 0:00:44.120 --> 0:00:49.720 protective stop. That was quick. Yeah, that's because I'm ran 0:00:49.800 --> 0:00:53.040 into the clothing. Yeah, you just basically impaled a stack 0:00:53.080 --> 0:00:55.480 of clothing. I always knew this game is big choice. 0:00:56.040 --> 0:00:58.440 But Chris Hayes, who's the guy with me here, has 0:00:58.480 --> 0:01:00.880 never heard that name. I'm not recognizing the name of 0:01:03.880 --> 0:01:08.960 glass In case, like a glass cub with a bunch 0:01:09.000 --> 0:01:11.319 of stuffed animals in the bottom, and then you have 0:01:11.400 --> 0:01:13.640 like two buttons and there's like a little river, and 0:01:13.800 --> 0:01:17.080 you like, oh, it's just like crane games. You could 0:01:17.440 --> 0:01:20.039 see it as an ultimate like virtual crane game in 0:01:20.080 --> 0:01:23.480 which you are trying well, it's not stacked against you, 0:01:23.920 --> 0:01:25.920 and which is stacked in your favor to try and 0:01:25.959 --> 0:01:28.800 help you actually grab something instead of drop everything you 0:01:28.880 --> 0:01:32.160 try to grab. My wife's actually really good at crane games. 0:01:32.200 --> 0:01:34.880 One's stuff from them all the time. Chris works for 0:01:34.959 --> 0:01:38.120 a startup called Kindred, which isn't quite in the business 0:01:38.120 --> 0:01:41.600 of producing nostalgically themed video games, and the game Chris 0:01:41.640 --> 0:01:43.640 is guiding me through isn't actually a game at all. 0:01:44.160 --> 0:01:46.840 It's the job of a dozen or so employees at Kindred. 0:01:47.560 --> 0:01:50.080 As they steer these robotic clause in a computer screen, 0:01:50.440 --> 0:01:52.880 they're teaching robots how to do things in the physical world. 0:01:53.520 --> 0:01:57.680 Kindred calls these employees pilots. Grabbing different objects and moving 0:01:57.680 --> 0:02:00.560 them around might sound like a simple task, but it's 0:02:00.560 --> 0:02:03.200 actually a lot harder than what most robots have been 0:02:03.280 --> 0:02:06.080 able to handle in the past. It requires them to 0:02:06.160 --> 0:02:10.359 constantly take in new information about their environments and adjust accordingly. 0:02:11.120 --> 0:02:13.560 The robots that can do this are just now emerging, 0:02:14.240 --> 0:02:16.639 and the better they get the more uncomfortable they make us. 0:02:17.000 --> 0:02:18.840 How many of our jobs are going to become obsolete? 0:02:19.600 --> 0:02:22.720 The short answer is, we don't know, but we do 0:02:22.840 --> 0:02:25.840 know something more about a related question, how are these 0:02:25.880 --> 0:02:42.919 robots going to learn to do our jobs? Anyway? Hi, 0:02:43.120 --> 0:02:47.160 I'm Akio and I'm Joshua Freusting, and this week on Decrypted, 0:02:47.280 --> 0:02:50.079 we're taking you inside the race to teach robots how 0:02:50.120 --> 0:02:52.400 to do the things that only humans could do before. 0:02:53.160 --> 0:02:55.359 Some of these robot trainers work at startups trying to 0:02:55.400 --> 0:02:58.640 sell robotics services to companies like The Gap. Some spend 0:02:58.639 --> 0:03:02.320 their time in academic slobs. Some do both. There's something 0:03:02.400 --> 0:03:05.560 ironic and bitter sweet about these people who are automating 0:03:05.600 --> 0:03:09.560 away their own usefulness. How they're approaching their work today 0:03:09.800 --> 0:03:11.840 might have lessons for all of us in the future. 0:03:12.360 --> 0:03:20.760 Stay with us. I walked right by the first time. 0:03:21.200 --> 0:03:25.160 There is not a sign on this door. Back in March, 0:03:25.240 --> 0:03:28.520 I visited Kindred's California office. It's in one of those 0:03:28.600 --> 0:03:31.679 charming San Francisco neighborhoods, you know, the ones under the 0:03:31.760 --> 0:03:34.160 highway over passes where you can't hear yourself think. And 0:03:34.200 --> 0:03:37.200 most of the businesses are auto repair shops. Hey, are 0:03:37.200 --> 0:03:42.680 you guys with Kindred? Hey, I'm looking for Chris some Josh. 0:03:43.040 --> 0:03:44.880 Chris is the guy who heard me talking to earlier. 0:03:45.320 --> 0:03:48.080 He was the very first pilot that Kindred hired. When 0:03:48.120 --> 0:03:50.000 I found him, he was sitting in a conference room 0:03:50.040 --> 0:03:52.200 on his laptop. The only thing that looked different from 0:03:52.200 --> 0:03:54.280 the setup that I used for my job was that 0:03:54.320 --> 0:03:56.520 he had two mice. There was the standard one your 0:03:56.600 --> 0:03:59.200 right hand, and a weird looking one in the left 0:03:59.240 --> 0:04:02.360 hand that he called us Space Navigator. He used both 0:04:02.440 --> 0:04:05.480 mice simultaneously to control a robot that the company had 0:04:05.520 --> 0:04:07.839 set up in its basement. So what did you see 0:04:07.880 --> 0:04:10.600 on his laptop? There were a few different video feeds 0:04:10.640 --> 0:04:13.240 of a pile of clothing laid out under the robotic claw. 0:04:13.680 --> 0:04:16.000 One was a standard video, the other one was what 0:04:16.080 --> 0:04:18.640 Chris called a polygon cloud, which was a three D 0:04:18.720 --> 0:04:20.800 image that showed the same thing but gave you a 0:04:20.800 --> 0:04:24.360 sense of depth. And a dashboard tracked your stats along 0:04:24.400 --> 0:04:26.280 the sides. You can tell how quickly we're picking things 0:04:26.360 --> 0:04:29.279 up on average, and things like that. And you said, 0:04:29.320 --> 0:04:32.479 there's an actual robot in the basement that was responding 0:04:32.480 --> 0:04:35.839 to his commands. Yeah, Kindred has a little workshop down 0:04:35.839 --> 0:04:38.320 there in the basement with a few engineers and the 0:04:38.480 --> 0:04:42.080 robot that it calls Kindred Sort. It's a yellow arm 0:04:42.160 --> 0:04:43.760 with a claw at the end, and it's an a 0:04:43.800 --> 0:04:47.679 round metal enclosure. The office has a few bins filled 0:04:47.680 --> 0:04:51.040 with stuff. It looks like someone just wandered around Kmart 0:04:51.200 --> 0:04:53.760 with a shopping cart, throwing random things in as they 0:04:53.760 --> 0:04:58.000 happened to walk past them. Protective stop. Just clothes and stuff. 0:04:58.720 --> 0:05:02.840 Somebody's in here. Um, let's see if they're the easier 0:05:02.880 --> 0:05:06.440 thing to try and grab you. We can try doing 0:05:06.480 --> 0:05:09.320 some training and if you feel like that's easy, we 0:05:09.360 --> 0:05:13.120 can throw some general merchandise in there as well. Okay, 0:05:13.440 --> 0:05:16.120 Then we walked back upstairs to Chris's computer. I sat 0:05:16.160 --> 0:05:18.280 in front of a program that showed several feeds from 0:05:18.279 --> 0:05:22.200 downstairs and tried to follow Chris's directions to move things around. 0:05:23.000 --> 0:05:25.240 It felt like a video game. Chris actually used to 0:05:25.240 --> 0:05:27.680 be a game designer, I should say. Since I was 0:05:27.720 --> 0:05:30.120 a kid, one of my biggest fears has been playing 0:05:30.240 --> 0:05:31.920 video games in front of people I don't really know, 0:05:32.600 --> 0:05:34.640 and so I wanted Chris to tell me how hard 0:05:34.680 --> 0:05:36.560 it was going to be. There's a trend in gaming 0:05:36.640 --> 0:05:38.920 that you want to make your game like easy to 0:05:38.960 --> 0:05:42.320 pick up, difficult to master, right, And I'd say it 0:05:42.480 --> 0:05:45.480 follows that pattern. It seems like it takes longer than 0:05:45.520 --> 0:05:48.839 twenty minutes to be competent at this, or at least 0:05:49.200 --> 0:05:52.360 I certainly wasn't teaching any robots anything of value after 0:05:52.400 --> 0:05:54.520 my twenty minutes test drive, so it feels like I 0:05:54.520 --> 0:05:56.799 could probably just grip it here you can try. Nope, 0:05:57.400 --> 0:05:59.040 there you go. There was your your claw game that 0:05:59.080 --> 0:06:02.640 felt very much like the log game. Actually, that sense 0:06:02.680 --> 0:06:06.960 of humiliation and disappointment I remember that, Josh, you have 0:06:07.040 --> 0:06:09.800 to ask, are you just a really bad gamer? So 0:06:09.880 --> 0:06:12.440 tell me straight, I'm not very good at this, not yet, 0:06:14.120 --> 0:06:16.520 but like on the average of someone who's spent fifteen 0:06:16.680 --> 0:06:20.760 minutes trying to standard Chris told me most people reach 0:06:20.800 --> 0:06:23.040 competence in a few hours, and then they can get 0:06:23.040 --> 0:06:25.560 about twice that good over a few days or whatever. 0:06:25.920 --> 0:06:28.480 And what do they do once they're good enough? They'll 0:06:28.480 --> 0:06:31.159 be in charge of driving robots operating in some distant 0:06:31.200 --> 0:06:34.400 warehouse somewhere. Kindred as a handful of clients. So far, 0:06:34.839 --> 0:06:37.160 the only one that's announced publicly is the Gap, which 0:06:37.240 --> 0:06:39.320 has a few of its robots running at a facility 0:06:39.320 --> 0:06:43.320 in Tennessee. Over time, as they learned from making my pilots, 0:06:43.880 --> 0:06:46.599 I would imagine that these robots become more capable of 0:06:46.640 --> 0:06:49.720 doing things on their own. That's the idea to teach 0:06:49.720 --> 0:06:52.479 a robot a task. The pilot starts off by mostly 0:06:52.520 --> 0:06:55.720 doing it himself. Then over time the robot takes over 0:06:56.000 --> 0:06:58.760 and just asks the pilot when they get stuck. Do 0:06:58.839 --> 0:07:01.200 we know how long it takes for them to reach 0:07:01.240 --> 0:07:06.200 this fully autonomous state where the robotsn't longer need the pilots? Now? 0:07:06.240 --> 0:07:08.800 This is a sense of question, of course, and robotics 0:07:08.839 --> 0:07:11.920 companies always tie themselves into pretzels trying to talk about 0:07:11.920 --> 0:07:15.320 their impact on labor. On the one hand, they want 0:07:15.320 --> 0:07:17.920 to boast about how quickly they're making progress, but they'll 0:07:17.920 --> 0:07:21.000 also object to any suggestion that they're quote unquote replacing 0:07:21.120 --> 0:07:24.520 jobs in these places. Here's George Bebow, a co founder 0:07:24.560 --> 0:07:28.640 of Kindred who's now the company's chief product officer. Jobs 0:07:28.640 --> 0:07:31.840 are going away. Nobody's seeing or talking about the jobs 0:07:31.840 --> 0:07:35.920 that are coming right, the annotators, the pilots and supervisors, 0:07:35.960 --> 0:07:38.640 the designers. There's a lot of work to be done, 0:07:39.400 --> 0:07:42.240 you know. There have been some pretty scary projections about 0:07:42.360 --> 0:07:46.040 job blasts. I'm thinking in particular about a McKinsey report 0:07:46.120 --> 0:07:49.520 from last year that predicted that automation could displace as 0:07:49.560 --> 0:07:53.840 many as an eight hundred million jobs by twenty Yeah, 0:07:53.920 --> 0:07:57.000 there's been a lot of bracing reports along those lines. Now, 0:07:57.040 --> 0:07:59.440 those numbers always do come with a lot of caveats. 0:08:00.040 --> 0:08:02.760 There could be fewer jobs that end up being automated away, 0:08:03.160 --> 0:08:05.360 and then there's also the question of how many new 0:08:05.480 --> 0:08:08.080 jobs could be created. And that same report that you cited, 0:08:08.400 --> 0:08:12.200 mackenzie said the newly created jobs might actually offset job losses. 0:08:12.760 --> 0:08:16.200 So the transition towards more automation could be disastrous. It 0:08:16.240 --> 0:08:19.000 could end up being much more benign, or it could 0:08:19.080 --> 0:08:21.560 end up being traumatic for many people, with the real 0:08:21.720 --> 0:08:24.840 issue being how to prepare workers for that shift. We 0:08:24.960 --> 0:08:31.560 just don't know. So it's being a robot pilot a job, 0:08:32.240 --> 0:08:34.440 you know. I asked everyone that kin Dre that question, 0:08:34.600 --> 0:08:36.920 and I have to say they didn't even steam sure 0:08:36.920 --> 0:08:40.000 how to answer it. As of now, the company doesn't 0:08:40.040 --> 0:08:43.320 actually have people who show up in the morning, drive 0:08:43.320 --> 0:08:47.199 a robot all day, and then head home at night. Instead, 0:08:47.559 --> 0:08:50.200 it's pilots, pilot for part of the day, then do 0:08:50.320 --> 0:08:53.120 something else for the rest of the time. Chris, their 0:08:53.160 --> 0:08:55.160 first pilot, said he's never done it for more than 0:08:55.200 --> 0:08:58.800 two hours at a stretch, and today he now spends 0:08:58.880 --> 0:09:01.320 his time working on the off where that other pilots use, 0:09:02.000 --> 0:09:04.800 So he's moved out of the robot piloting job altogether. 0:09:06.160 --> 0:09:08.760 Sounds like a smart move if you're kind of automating 0:09:08.800 --> 0:09:12.160 yourself out of a job anyway. Yeah, Well, the folks 0:09:12.160 --> 0:09:15.720 that kindred say the demand for pilots will actually grow significantly. 0:09:16.480 --> 0:09:19.480 That's because the number of robots in the field should 0:09:19.559 --> 0:09:22.680 grow faster than the rate at which each robot becomes 0:09:22.720 --> 0:09:27.120 completely autonomous. So I'm imagining these pilots maybe controlling twenty 0:09:27.200 --> 0:09:30.440 robots at a time instead of maybe just one or two. Yeah, 0:09:30.480 --> 0:09:32.520 that's where they want to end up. And one of 0:09:32.520 --> 0:09:36.040 the challenges Chris is dealing with now actually relates to 0:09:36.120 --> 0:09:38.080 what it's like for a single pilot to be in 0:09:38.160 --> 0:09:40.880 charge of so many robots. He says it can be 0:09:40.960 --> 0:09:43.959 jarring to finish helping one machine pick up a shirt 0:09:44.320 --> 0:09:47.200 and then suddenly be transported to another robot that's right 0:09:47.200 --> 0:09:49.200 in the middle of trying to do something else. The 0:09:49.240 --> 0:09:53.240 more machines you switch between, the more I guess that 0:09:53.559 --> 0:09:56.520 kind of impacts you mentally. Is far as like contact switching. 0:09:56.960 --> 0:09:59.400 When you have you switch over to another machine, you 0:09:59.600 --> 0:10:01.920 it takes little bit of time to get the context 0:10:02.040 --> 0:10:05.120 of like what's going on here, like what is the 0:10:05.120 --> 0:10:07.439 best I'm to grab and like go for it right, 0:10:08.800 --> 0:10:12.839 And the more different machines you have, the more that 0:10:12.920 --> 0:10:18.000 kind of impacts things. Okay, before I described my next stop, 0:10:18.120 --> 0:10:19.720 I think it would be helpful to run through some 0:10:19.760 --> 0:10:23.480 basic FOCAB. So let's start with what we've described so far. 0:10:24.320 --> 0:10:27.520 When I was driving that claw at Kindred, the robotic 0:10:27.600 --> 0:10:31.800 system was tracking my movements and learning exactly how I 0:10:31.920 --> 0:10:36.880 picked up a shirt. That technique is called imitation learning. Okay, 0:10:36.920 --> 0:10:39.440 so this is like showing your kids how to tie 0:10:39.480 --> 0:10:44.240 their shoes by having them watch you do it first exactly. Now, 0:10:44.240 --> 0:10:48.160 there's another strategy for teaching robots that's called reinforcement learning. 0:10:48.679 --> 0:10:50.560 Let's say you're trying to teach a robot how to 0:10:50.800 --> 0:10:54.120 do a backflip. It gives it a shot, but a 0:10:54.200 --> 0:10:56.880 robot doesn't even know what a backflip is, so you 0:10:56.920 --> 0:10:59.600 have to tell it whether it's successful or not. You'd 0:10:59.600 --> 0:11:01.680 also tell at how close it got to doing it right, 0:11:02.160 --> 0:11:05.200 and maybe just how it went wrong. Then the robot 0:11:05.240 --> 0:11:08.840 tries another backflip, and another one and another one, and 0:11:08.880 --> 0:11:11.600 through trial and error, it learns exactly what a backflip 0:11:11.679 --> 0:11:15.120 is and how to successfully carry one out. So this 0:11:15.200 --> 0:11:17.600 might be more like giving your kids some candy when 0:11:17.600 --> 0:11:21.280 they do their homework on time exactly, And just like 0:11:21.440 --> 0:11:24.640 with teaching children, when you're teaching robots, it can help 0:11:24.679 --> 0:11:29.320 to use a combination of different teaching techniques. Reinforcement learning 0:11:29.320 --> 0:11:31.560 has been the dominant technique and a lot of artificial 0:11:31.559 --> 0:11:34.800 intelligence research in recent years, and it works best when 0:11:34.840 --> 0:11:37.560 you have a huge amount of data, but imitation learning 0:11:37.600 --> 0:11:40.640 can fill in some key gaps, so places like Kindred 0:11:40.679 --> 0:11:43.560 that are focusing on it use a combination of both techniques. 0:11:45.280 --> 0:11:48.079 Some of the most advanced academic research on these techniques 0:11:48.360 --> 0:11:51.040 is going on just across the bay from Kindred at 0:11:51.040 --> 0:11:58.319 the Robotics Lab at Berkeley h I was there to 0:11:58.360 --> 0:12:01.800 meet Chelsea Finn. She's a pH d student there. Chelsea 0:12:01.880 --> 0:12:04.080 is a quiet, friendly type who was happy to walk 0:12:04.080 --> 0:12:05.480 me through what she did and was really good at 0:12:05.480 --> 0:12:08.200 explaining things. But I also got the feeling that she 0:12:08.320 --> 0:12:12.000 was a bit embarrassed by the attention. Chelsea's research is 0:12:12.040 --> 0:12:15.240 kind of like the graduate level coursework for robots. The 0:12:15.280 --> 0:12:18.400 ideas that robots should be able to learn things progressively 0:12:18.480 --> 0:12:21.280 faster than the last ones by building on the knowledge 0:12:21.280 --> 0:12:23.000 that they've picked up from all the other things they've 0:12:23.080 --> 0:12:25.439 learned to do well. We've been building on techniques which 0:12:25.480 --> 0:12:27.400 are called metal learning or learning to learn, so that 0:12:27.440 --> 0:12:30.600 you can learn from a new demonstration with a very 0:12:30.600 --> 0:12:33.040 small amount of data. When you see a new demonstration, 0:12:33.679 --> 0:12:36.080 a single new demonstration for a new task, you can 0:12:36.160 --> 0:12:39.120 learn that just from that task. A big goal for 0:12:39.200 --> 0:12:42.040 Chelsea and her colleagues is to have robots learned from 0:12:42.200 --> 0:12:46.600 less and less exact demonstrations. Give us an example. Okay, 0:12:46.600 --> 0:12:49.080 this is gonna sound very elementary about human standards, but 0:12:49.120 --> 0:12:51.880 here goes. Let's say you have a robot and you 0:12:51.880 --> 0:12:53.880 wanted to pick up a ball and put it in 0:12:53.920 --> 0:12:56.680 a red bowl. You've already taught this robot how to 0:12:56.679 --> 0:12:59.400 pick things up, and you've taught it how to put 0:12:59.400 --> 0:13:03.000 things two containers, but it's never seen a red bull before. 0:13:04.160 --> 0:13:07.120 So pilots like Chris and the folks at Kindred would 0:13:07.120 --> 0:13:09.640 do this in a direct way. They'd steer the robot 0:13:09.679 --> 0:13:13.120 through the task while the censors record exactly what this means. 0:13:13.960 --> 0:13:16.679 But Chelsea's team would just showed a video of someone 0:13:16.720 --> 0:13:20.080 putting a ball in a red bowl. At first, the 0:13:20.080 --> 0:13:22.120 person would have to be in exactly the same environment 0:13:22.120 --> 0:13:25.800 the robot already knew about same background, same table, same ball. 0:13:26.480 --> 0:13:28.640 But over time it had learned to do things from 0:13:28.720 --> 0:13:31.400 videos that were less and less similar to its own environment, 0:13:32.040 --> 0:13:34.960 and eventually robots were able to pick up new skills 0:13:35.080 --> 0:13:37.880 just by searching for videos of people, say, putting balls 0:13:37.920 --> 0:13:40.760 into red bulls. It's a big genre on YouTube, and 0:13:41.360 --> 0:13:44.680 the endgame would be for robots to just search YouTube 0:13:44.720 --> 0:13:47.240 for videos it needed, just the way that you might 0:13:47.320 --> 0:13:49.640 learn how to braid hair or change a light switch. 0:13:50.679 --> 0:13:55.240 I'm imagining Chelsea's robot watching hours of makeup tutorials. But 0:13:55.360 --> 0:13:58.040 Chelsea and her colleagues are slowly working their way out 0:13:58.120 --> 0:14:00.880 of the lab. At a big machine and learning expo 0:14:01.000 --> 0:14:04.600 called the Conference on Neural Information Processing Systems, or NIPS, 0:14:05.000 --> 0:14:08.040 they recently showed off the robots flexibility by having people 0:14:08.040 --> 0:14:11.080 at the conference record themselves doing some simple task a 0:14:11.120 --> 0:14:13.920 single time, and then showing how the robot could do 0:14:13.960 --> 0:14:16.000 the same thing by watching them. I mean, this is 0:14:16.080 --> 0:14:18.480 kind of a step between the lab and the real world, 0:14:18.520 --> 0:14:20.480 and that you're actually moving off of the lab that 0:14:20.480 --> 0:14:24.800 it was trained on into this other space and isn't 0:14:25.200 --> 0:14:27.480 the real world. No, it's it's not not. Yes, it's 0:14:27.480 --> 0:14:30.200 definitely not the real world. Although I will say that 0:14:30.240 --> 0:14:32.520 the researchers really like to try to trick the robot. 0:14:33.840 --> 0:14:35.960 You know, I can see how this would really speed 0:14:35.960 --> 0:14:38.200 things up. If you didn't have to have someone like 0:14:38.280 --> 0:14:41.960 Chris painstakingly show how to do each and everything, these 0:14:42.040 --> 0:14:44.720 robots could learn a lot faster. Yeah, And the point 0:14:44.720 --> 0:14:46.840 of all this is that robots don't have to do 0:14:46.920 --> 0:14:49.280 just a single type of job. You have to be 0:14:49.320 --> 0:14:51.680 prepared to do all kinds of things depending on what 0:14:51.720 --> 0:14:54.560 people need from them at any specific time, which is 0:14:54.640 --> 0:14:57.160 probably what most of our jobs are like. Right. Think 0:14:57.200 --> 0:15:00.240 about what makes a good human colleague. It's based really 0:15:00.320 --> 0:15:02.520 the ones who are the most flexible at doing all 0:15:02.600 --> 0:15:05.240 kinds of things and the ones that pick up things 0:15:05.280 --> 0:15:09.160 with these amount of handholding. Right. But Chelsea thinks there's 0:15:09.160 --> 0:15:11.680 a long way to go before robots are really able 0:15:11.680 --> 0:15:14.120 to be this kind of generalist usually a long way. 0:15:14.200 --> 0:15:16.480 Is it like it's so far in the future that 0:15:16.520 --> 0:15:19.240 it's hard to even imagine? Is that like five years, 0:15:19.440 --> 0:15:22.720 that ten years? I would say that it's more than 0:15:22.800 --> 0:15:26.120 five years, and that beyond five years, it's uh, it's 0:15:26.160 --> 0:15:32.200 hard to make accurate predictions. Right down the hall from 0:15:32.240 --> 0:15:34.520 Chelsea is someone who's trying to bridge the gap between 0:15:34.520 --> 0:15:37.600 the academic and commercial worlds of robotics. His name is 0:15:37.640 --> 0:15:41.280 Peter Abiel. He recently started a company called Embodied Intelligence, 0:15:41.320 --> 0:15:43.480 but keeps an opposit at Berkeley. So I swung by 0:15:43.520 --> 0:15:46.400 after saying goodbye to Chelsea. Peter just returned from a 0:15:46.440 --> 0:15:49.440 private event that Jeff Bezos throws every year for researchers 0:15:49.440 --> 0:15:53.200 and AI, robotics and space so they can compare notes. Yeah, 0:15:53.200 --> 0:15:55.920 I've read about the party's I think Dasis calls them 0:15:55.960 --> 0:15:59.680 summer camp for geeks. Well, Peter definitely meets that description. 0:16:00.240 --> 0:16:02.880 When I found him, he was unwrapping random pieces of 0:16:02.920 --> 0:16:04.960 computer equipment that had showed up in the mail while 0:16:04.960 --> 0:16:09.200 he was away. Peter thinks the most important thing to 0:16:09.200 --> 0:16:12.480 make progress on is accelerating the pace at which robots 0:16:12.480 --> 0:16:15.360 to learn new tasks. Hundred and first skills should be 0:16:15.400 --> 0:16:19.080 faster than the hundred skill to learn, and then as 0:16:19.080 --> 0:16:21.240 you keep going, the next one should be even faster 0:16:21.360 --> 0:16:23.800 than faster, even faster. And that's where things get very interesting. 0:16:23.840 --> 0:16:26.160 Is at some point is going to hit a very 0:16:26.200 --> 0:16:29.080 low number of demonstrations, maybe only one or two to 0:16:29.200 --> 0:16:33.280 learn something new. Peter's business, embodied intelligence is super early. 0:16:33.800 --> 0:16:36.080 At this point. He isn't willing to reveal much about 0:16:36.080 --> 0:16:38.480 it except that it exists and he's been talking to 0:16:38.560 --> 0:16:42.080 potential clients. Some of his more recent work at Berkeley 0:16:42.280 --> 0:16:45.880 focused on robot pilots who wore virtual reality headsets. The 0:16:45.920 --> 0:16:48.480 researchers were surprised that it took less than thirty minutes 0:16:48.520 --> 0:16:51.520 of demonstration for robots to learn most new tasks. So 0:16:51.680 --> 0:16:54.400 Josh is he in Chelsea's camp that it will take 0:16:54.560 --> 0:16:57.120 at least five years for robots to be able to 0:16:57.120 --> 0:17:00.400 handle the wide range of tasks that we do well. 0:17:00.440 --> 0:17:02.840 I think the main difference between Peter and Chelsea is 0:17:03.000 --> 0:17:06.200 almost want of tone. Chelsea is speaking as the cautious 0:17:06.200 --> 0:17:10.760 academic and Peter is more of the excitable entrepreneur. They 0:17:10.800 --> 0:17:13.880 both agree that robots that are just sort of generally 0:17:13.920 --> 0:17:16.879 competent and everything that's more than five years out and 0:17:17.080 --> 0:17:20.040 kind of hard to tell. But Peter focus is more 0:17:20.240 --> 0:17:22.359 on what they will be able to do in five years, 0:17:22.400 --> 0:17:24.280 and he thinks they'll be good enough to try a 0:17:24.320 --> 0:17:26.919 lot of things that they can't do today. The longer 0:17:26.960 --> 0:17:29.520 timeline is can we get anybody to teach a robot 0:17:29.560 --> 0:17:33.240 and new skill? The shorter timeline is, can we ourselves 0:17:33.520 --> 0:17:37.720 at embodied intelligence teach robots new skills that robots haven't 0:17:37.720 --> 0:17:39.960 done before. So if you can do that, then those 0:17:40.040 --> 0:17:42.680 robots could be doing things that currently are not possible. 0:17:43.160 --> 0:17:45.239 You know. It sounds like the speed that robots can 0:17:45.359 --> 0:17:48.080 learn these new tasks and switch from task to task 0:17:48.600 --> 0:17:50.920 could be a really critical point to how they would 0:17:50.960 --> 0:17:54.199 be used. Yeah, take the example of a company that 0:17:54.320 --> 0:17:58.040 runs huge logistics facilities. They need a lot of robots 0:17:58.080 --> 0:18:01.120 to operate really quickly, But what they're gonna do from 0:18:01.200 --> 0:18:04.120 day to day isn't going to change that much. If 0:18:04.119 --> 0:18:07.040 it's an expensive and time consuming process for teaching those 0:18:07.119 --> 0:18:09.399 robots their jobs in the first place, it might not 0:18:09.480 --> 0:18:11.960 matter that much because once they're set up they can 0:18:11.960 --> 0:18:14.680 be very efficient. But I would imagine at a smaller 0:18:14.720 --> 0:18:18.040 business the needs would change a lot more than a 0:18:18.119 --> 0:18:22.000 big industrialized setting. Yeah, the way that smaller operations are 0:18:22.000 --> 0:18:24.919 going to benefit from robots is probably if they can 0:18:24.960 --> 0:18:28.119 be really customizable and if they can be trained quickly 0:18:28.160 --> 0:18:30.640 to do whatever the task is at hand. And that's 0:18:30.640 --> 0:18:32.440 really where Peter is hoping to get to. He wants 0:18:32.480 --> 0:18:35.000 to change the dynamic of what it means to have 0:18:35.119 --> 0:18:37.280 a robot. Do Do I need help? Do I need 0:18:37.320 --> 0:18:39.879 physical help moving things around? Putting things together? Because if 0:18:39.920 --> 0:18:42.840 I do whatever it is I need today, it's fun. 0:18:42.920 --> 0:18:44.800 I can need something else tomorrow. I can just reteach 0:18:44.840 --> 0:18:47.320 the robot every single time, and so that will open 0:18:47.400 --> 0:18:50.439 up a lot of opportunities and people will need far 0:18:50.520 --> 0:18:53.600 less capital investment, far less scale be able to take 0:18:53.640 --> 0:18:57.920 advantage of having a robot helped them out. Okay, going 0:18:57.960 --> 0:19:00.600 back to Kindred, the job of a rope but pilot 0:19:00.800 --> 0:19:03.399 didn't actually always look like what I tried sitting in 0:19:03.480 --> 0:19:05.639 front of a laptop and playing what seemed like a 0:19:05.680 --> 0:19:09.199 pretty simple video game. In fact, the original mission of 0:19:09.240 --> 0:19:12.200 the company was to create robots that could do anything 0:19:12.240 --> 0:19:15.760 a human could do physically. At least according to George, 0:19:15.800 --> 0:19:18.399 the co founder who we talked earlier, the job of 0:19:18.480 --> 0:19:21.560 robot training used to involve wearing what's called an exoskeleton. 0:19:22.160 --> 0:19:24.000 Is that what you can capture all the motions from 0:19:24.000 --> 0:19:26.919 all the joints, all the fingers, elbose shoulders, you can 0:19:26.960 --> 0:19:29.280 capture all the informations basically a person in a robots 0:19:29.280 --> 0:19:31.960 the person in robots. For listeners who have never seen 0:19:32.000 --> 0:19:36.480 an exo skeleton, they look truly crazy, like Iron Man. 0:19:36.640 --> 0:19:38.840 It really looks like you're wearing a robot on your body. 0:19:39.160 --> 0:19:41.640 And this is really exciting for scientists. But it turns 0:19:41.640 --> 0:19:44.520 out it was overkill for what most of Kindred's potential 0:19:44.560 --> 0:19:48.080 customers said they wanted, like placing boxes on their right 0:19:48.119 --> 0:19:51.920 shelves right. So the company set the exo skeletons aside 0:19:52.440 --> 0:19:56.320 and focused on doing this one rather mundane thing that 0:19:56.359 --> 0:19:59.320 would win them contracts with companies that were operating warehouses. 0:20:00.080 --> 0:20:02.080 That does feel like a little bit of a letdown. 0:20:02.560 --> 0:20:05.600 The company actually split over it. Two of its founders 0:20:05.720 --> 0:20:07.520 left at the beginning of this year to start a 0:20:07.520 --> 0:20:11.320 new company called Sanctuary dot ai. I talked to one 0:20:11.320 --> 0:20:14.600 of them. Her name is Suzanne Gildert, and she said 0:20:14.680 --> 0:20:16.960 that she had wanted to do work on training robots 0:20:17.000 --> 0:20:20.159 to get all the way to human like intelligence. But 0:20:20.240 --> 0:20:21.639 she also knew that this was going to take a 0:20:21.640 --> 0:20:24.320 lot of time toiling without any real business model in 0:20:24.320 --> 0:20:27.560 the short term, and when Kindred discovered something it could 0:20:27.560 --> 0:20:31.199 sell immediately, she decided to part ways. And is it 0:20:31.280 --> 0:20:34.359 selling well? Kindred does say it is, but they're being 0:20:34.359 --> 0:20:36.960 pretty vague about it. The company's testing period at the 0:20:37.000 --> 0:20:40.600 Gap ended earlier this year, and they decided to expand it. 0:20:41.000 --> 0:20:43.800 I caught up with Jim Liefer, Kindred CEO, on the 0:20:43.800 --> 0:20:46.399 phone a few weeks ago. He was at a supply 0:20:46.480 --> 0:20:49.240 chain conference in Atlanta and said his voice was sore 0:20:49.280 --> 0:20:51.680 from all the pitching. We have a very full dance 0:20:51.760 --> 0:20:55.320 card after these four days. I'm pretty excited about it. 0:20:55.600 --> 0:20:57.560 Jim says that Kindred is really going to ramp up 0:20:57.600 --> 0:21:00.879 its operations over the course of and will operate on 0:21:00.920 --> 0:21:03.920