The Race to Teach Robots How to Do Our Jobs

was a kid. I'm sitting in front of a laptop, using a mouse in each hand to steer a claw around my screen and trying to grab plush objects. This is where your pointer is. Make sure you're not clicking buttons and so what you've just done is caused a protective stop. That was quick. Yeah, that's because I'm ran into the clothing. Yeah, you just basically impaled a stack of clothing. I always knew this game is big choice. But Chris Hayes, who's the guy with me here, has

never heard that name. I'm not recognizing the name of glass In case, like a glass cub with a bunch of stuffed animals in the bottom, and then you have like two buttons and there's like a little river, and you like, oh, it's just like crane games. You could see it as an ultimate like virtual crane game in which you are trying well, it's not stacked against you, and which is stacked in your favor to try and help you actually grab something instead of drop everything you

try to grab. My wife's actually really good at crane games. One's stuff from them all the time. Chris works for a startup called Kindred, which isn't quite in the business of producing nostalgically themed video games, and the game Chris is guiding me through isn't actually a game at all. It's the job of a dozen or so employees at Kindred. As they steer these robotic clause in a computer screen, they're teaching robots how to do things in the physical world.

Kindred calls these employees pilots. Grabbing different objects and moving them around might sound like a simple task, but it's actually a lot harder than what most robots have been able to handle in the past. It requires them to constantly take in new information about their environments and adjust accordingly. The robots that can do this are just now emerging, and the better they get the more uncomfortable they make us. How many of our jobs are going to become obsolete?

The short answer is, we don't know, but we do know something more about a related question, how are these robots going to learn to do our jobs? Anyway? Hi, I'm Akio and I'm Joshua Freusting, and this week on Decrypted, we're taking you inside the race to teach robots how to do the things that only humans could do before. Some of these robot trainers work at startups trying to sell robotics services to companies like The Gap. Some spend

their time in academic slobs. Some do both. There's something ironic and bitter sweet about these people who are automating away their own usefulness. How they're approaching their work today might have lessons for all of us in the future. Stay with us. I walked right by the first time. There is not a sign on this door. Back in March, I visited Kindred's California office. It's in one of those charming San Francisco neighborhoods, you know, the ones under the

highway over passes where you can't hear yourself think. And most of the businesses are auto repair shops. Hey, are you guys with Kindred? Hey, I'm looking for Chris some Josh. Chris is the guy who heard me talking to earlier. He was the very first pilot that Kindred hired. When I found him, he was sitting in a conference room on his laptop. The only thing that looked different from the setup that I used for my job was that

he had two mice. There was the standard one your right hand, and a weird looking one in the left hand that he called us Space Navigator. He used both mice simultaneously to control a robot that the company had set up in its basement. So what did you see on his laptop? There were a few different video feeds of a pile of clothing laid out under the robotic claw.

One was a standard video, the other one was what Chris called a polygon cloud, which was a three D image that showed the same thing but gave you a sense of depth. And a dashboard tracked your stats along the sides. You can tell how quickly we're picking things up on average, and things like that. And you said, there's an actual robot in the basement that was responding to his commands. Yeah, Kindred has a little workshop down there in the basement with a few engineers and the

robot that it calls Kindred Sort. It's a yellow arm with a claw at the end, and it's an a round metal enclosure. The office has a few bins filled with stuff. It looks like someone just wandered around Kmart with a shopping cart, throwing random things in as they happened to walk past them. Protective stop. Just clothes and stuff. Somebody's in here. Um, let's see if they're the easier thing to try and grab you. We can try doing some training and if you feel like that's easy, we

can throw some general merchandise in there as well. Okay, Then we walked back upstairs to Chris's computer. I sat in front of a program that showed several feeds from downstairs and tried to follow Chris's directions to move things around. It felt like a video game. Chris actually used to be a game designer, I should say. Since I was a kid, one of my biggest fears has been playing video games in front of people I don't really know, and so I wanted Chris to tell me how hard

it was going to be. There's a trend in gaming that you want to make your game like easy to pick up, difficult to master, right, And I'd say it follows that pattern. It seems like it takes longer than twenty minutes to be competent at this, or at least I certainly wasn't teaching any robots anything of value after my twenty minutes test drive, so it feels like I could probably just grip it here you can try. Nope, there you go. There was your your claw game that

felt very much like the log game. Actually, that sense of humiliation and disappointment I remember that, Josh, you have to ask, are you just a really bad gamer? So tell me straight, I'm not very good at this, not yet, but like on the average of someone who's spent fifteen minutes trying to standard Chris told me most people reach competence in a few hours, and then they can get about twice that good over a few days or whatever.

And what do they do once they're good enough? They'll be in charge of driving robots operating in some distant warehouse somewhere. Kindred as a handful of clients. So far, the only one that's announced publicly is the Gap, which has a few of its robots running at a facility in Tennessee. Over time, as they learned from making my pilots, I would imagine that these robots become more capable of doing things on their own. That's the idea to teach a robot a task. The pilot starts off by mostly

doing it himself. Then over time the robot takes over and just asks the pilot when they get stuck. Do we know how long it takes for them to reach this fully autonomous state where the robotsn't longer need the pilots? Now? This is a sense of question, of course, and robotics companies always tie themselves into pretzels trying to talk about

their impact on labor. On the one hand, they want to boast about how quickly they're making progress, but they'll also object to any suggestion that they're quote unquote replacing jobs in these places. Here's George Bebow, a co founder of Kindred who's now the company's chief product officer. Jobs are going away. Nobody's seeing or talking about the jobs that are coming right, the annotators, the pilots and supervisors, the designers. There's a lot of work to be done,

you know. There have been some pretty scary projections about job blasts. I'm thinking in particular about a McKinsey report from last year that predicted that automation could displace as many as an eight hundred million jobs by twenty Yeah, there's been a lot of bracing reports along those lines. Now, those numbers always do come with a lot of caveats. There could be fewer jobs that end up being automated away, and then there's also the question of how many new

jobs could be created. And that same report that you cited, mackenzie said the newly created jobs might actually offset job losses. So the transition towards more automation could be disastrous. It could end up being much more benign, or it could end up being traumatic for many people, with the real issue being how to prepare workers for that shift. We just don't know. So it's being a robot pilot a job,

you know. I asked everyone that kin Dre that question, and I have to say they didn't even steam sure how to answer it. As of now, the company doesn't actually have people who show up in the morning, drive a robot all day, and then head home at night. Instead, it's pilots, pilot for part of the day, then do

something else for the rest of the time. Chris, their first pilot, said he's never done it for more than two hours at a stretch, and today he now spends his time working on the off where that other pilots use, So he's moved out of the robot piloting job altogether. Sounds like a smart move if you're kind of automating yourself out of a job anyway. Yeah, Well, the folks that kindred say the demand for pilots will actually grow significantly.

That's because the number of robots in the field should grow faster than the rate at which each robot becomes completely autonomous. So I'm imagining these pilots maybe controlling twenty robots at a time instead of maybe just one or two. Yeah, that's where they want to end up. And one of the challenges Chris is dealing with now actually relates to what it's like for a single pilot to be in

charge of so many robots. He says it can be jarring to finish helping one machine pick up a shirt and then suddenly be transported to another robot that's right in the middle of trying to do something else. The more machines you switch between, the more I guess that kind of impacts you mentally. Is far as like contact switching.

When you have you switch over to another machine, you it takes little bit of time to get the context of like what's going on here, like what is the best I'm to grab and like go for it right, And the more different machines you have, the more that kind of impacts things. Okay, before I described my next stop, I think it would be helpful to run through some basic FOCAB. So let's start with what we've described so far.

When I was driving that claw at Kindred, the robotic system was tracking my movements and learning exactly how I picked up a shirt. That technique is called imitation learning. Okay, so this is like showing your kids how to tie their shoes by having them watch you do it first exactly. Now, there's another strategy for teaching robots that's called reinforcement learning. Let's say you're trying to teach a robot how to

do a backflip. It gives it a shot, but a robot doesn't even know what a backflip is, so you have to tell it whether it's successful or not. You'd also tell at how close it got to doing it right, and maybe just how it went wrong. Then the robot tries another backflip, and another one and another one, and through trial and error, it learns exactly what a backflip

is and how to successfully carry one out. So this might be more like giving your kids some candy when they do their homework on time exactly, And just like with teaching children, when you're teaching robots, it can help

to use a combination of different teaching techniques. Reinforcement learning has been the dominant technique and a lot of artificial intelligence research in recent years, and it works best when you have a huge amount of data, but imitation learning can fill in some key gaps, so places like Kindred that are focusing on it use a combination of both techniques.

Some of the most advanced academic research on these techniques is going on just across the bay from Kindred at the Robotics Lab at Berkeley h I was there to meet Chelsea Finn. She's a pH d student there. Chelsea is a quiet, friendly type who was happy to walk me through what she did and was really good at explaining things. But I also got the feeling that she was a bit embarrassed by the attention. Chelsea's research is

kind of like the graduate level coursework for robots. The ideas that robots should be able to learn things progressively faster than the last ones by building on the knowledge that they've picked up from all the other things they've learned to do well. We've been building on techniques which are called metal learning or learning to learn, so that you can learn from a new demonstration with a very

small amount of data. When you see a new demonstration, a single new demonstration for a new task, you can learn that just from that task. A big goal for Chelsea and her colleagues is to have robots learned from less and less exact demonstrations. Give us an example. Okay, this is gonna sound very elementary about human standards, but here goes. Let's say you have a robot and you wanted to pick up a ball and put it in

a red bowl. You've already taught this robot how to pick things up, and you've taught it how to put things two containers, but it's never seen a red bull before. So pilots like Chris and the folks at Kindred would do this in a direct way. They'd steer the robot through the task while the censors record exactly what this means. But Chelsea's team would just showed a video of someone

putting a ball in a red bowl. At first, the person would have to be in exactly the same environment the robot already knew about same background, same table, same ball. But over time it had learned to do things from videos that were less and less similar to its own environment, and eventually robots were able to pick up new skills just by searching for videos of people, say, putting balls

into red bulls. It's a big genre on YouTube, and the endgame would be for robots to just search YouTube for videos it needed, just the way that you might learn how to braid hair or change a light switch. I'm imagining Chelsea's robot watching hours of makeup tutorials. But Chelsea and her colleagues are slowly working their way out

of the lab. At a big machine and learning expo called the Conference on Neural Information Processing Systems, or NIPS, they recently showed off the robots flexibility by having people at the conference record themselves doing some simple task a single time, and then showing how the robot could do the same thing by watching them. I mean, this is kind of a step between the lab and the real world, and that you're actually moving off of the lab that it was trained on into this other space and isn't

the real world. No, it's it's not not. Yes, it's definitely not the real world. Although I will say that the researchers really like to try to trick the robot. You know, I can see how this would really speed things up. If you didn't have to have someone like Chris painstakingly show how to do each and everything, these robots could learn a lot faster. Yeah, And the point of all this is that robots don't have to do

just a single type of job. You have to be prepared to do all kinds of things depending on what people need from them at any specific time, which is probably what most of our jobs are like. Right. Think about what makes a good human colleague. It's based really the ones who are the most flexible at doing all kinds of things and the ones that pick up things

with these amount of handholding. Right. But Chelsea thinks there's a long way to go before robots are really able to be this kind of generalist usually a long way. Is it like it's so far in the future that it's hard to even imagine? Is that like five years, that ten years? I would say that it's more than five years, and that beyond five years, it's uh, it's hard to make accurate predictions. Right down the hall from Chelsea is someone who's trying to bridge the gap between

the academic and commercial worlds of robotics. His name is Peter Abiel. He recently started a company called Embodied Intelligence, but keeps an opposit at Berkeley. So I swung by after saying goodbye to Chelsea. Peter just returned from a private event that Jeff Bezos throws every year for researchers and AI, robotics and space so they can compare notes. Yeah, I've read about the party's I think Dasis calls them summer camp for geeks. Well, Peter definitely meets that description.

When I found him, he was unwrapping random pieces of computer equipment that had showed up in the mail while he was away. Peter thinks the most important thing to make progress on is accelerating the pace at which robots to learn new tasks. Hundred and first skills should be faster than the hundred skill to learn, and then as you keep going, the next one should be even faster than faster, even faster. And that's where things get very interesting.

Is at some point is going to hit a very low number of demonstrations, maybe only one or two to learn something new. Peter's business, embodied intelligence is super early. At this point. He isn't willing to reveal much about it except that it exists and he's been talking to potential clients. Some of his more recent work at Berkeley focused on robot pilots who wore virtual reality headsets. The researchers were surprised that it took less than thirty minutes

of demonstration for robots to learn most new tasks. So Josh is he in Chelsea's camp that it will take at least five years for robots to be able to handle the wide range of tasks that we do well. I think the main difference between Peter and Chelsea is almost want of tone. Chelsea is speaking as the cautious academic and Peter is more of the excitable entrepreneur. They both agree that robots that are just sort of generally competent and everything that's more than five years out and

kind of hard to tell. But Peter focus is more on what they will be able to do in five years, and he thinks they'll be good enough to try a lot of things that they can't do today. The longer timeline is can we get anybody to teach a robot and new skill? The shorter timeline is, can we ourselves at embodied intelligence teach robots new skills that robots haven't done before. So if you can do that, then those robots could be doing things that currently are not possible.

You know. It sounds like the speed that robots can learn these new tasks and switch from task to task could be a really critical point to how they would be used. Yeah, take the example of a company that runs huge logistics facilities. They need a lot of robots to operate really quickly, But what they're gonna do from

day to day isn't going to change that much. If it's an expensive and time consuming process for teaching those robots their jobs in the first place, it might not matter that much because once they're set up they can be very efficient. But I would imagine at a smaller business the needs would change a lot more than a big industrialized setting. Yeah, the way that smaller operations are going to benefit from robots is probably if they can be really customizable and if they can be trained quickly

to do whatever the task is at hand. And that's really where Peter is hoping to get to. He wants to change the dynamic of what it means to have a robot. Do Do I need help? Do I need physical help moving things around? Putting things together? Because if I do whatever it is I need today, it's fun.

I can need something else tomorrow. I can just reteach the robot every single time, and so that will open up a lot of opportunities and people will need far less capital investment, far less scale be able to take advantage of having a robot helped them out. Okay, going back to Kindred, the job of a rope but pilot didn't actually always look like what I tried sitting in front of a laptop and playing what seemed like a

pretty simple video game. In fact, the original mission of the company was to create robots that could do anything a human could do physically. At least according to George, the co founder who we talked earlier, the job of robot training used to involve wearing what's called an exoskeleton. Is that what you can capture all the motions from all the joints, all the fingers, elbose shoulders, you can capture all the informations basically a person in a robots

the person in robots. For listeners who have never seen an exo skeleton, they look truly crazy, like Iron Man. It really looks like you're wearing a robot on your body. And this is really exciting for scientists. But it turns out it was overkill for what most of Kindred's potential customers said they wanted, like placing boxes on their right

shelves right. So the company set the exo skeletons aside and focused on doing this one rather mundane thing that would win them contracts with companies that were operating warehouses. That does feel like a little bit of a letdown. The company actually split over it. Two of its founders left at the beginning of this year to start a new company called Sanctuary dot ai. I talked to one of them. Her name is Suzanne Gildert, and she said that she had wanted to do work on training robots

to get all the way to human like intelligence. But she also knew that this was going to take a lot of time toiling without any real business model in the short term, and when Kindred discovered something it could sell immediately, she decided to part ways. And is it selling well? Kindred does say it is, but they're being pretty vague about it. The company's testing period at the Gap ended earlier this year, and they decided to expand it. I caught up with Jim Liefer, Kindred CEO, on the

phone a few weeks ago. He was at a supply chain conference in Atlanta and said his voice was sore from all the pitching. We have a very full dance card after these four days. I'm pretty excited about it. Jim says that Kindred is really going to ramp up its operations over the course of and will operate on a far larger commercial scale. I wanted to know whether the company's view on the roll of the robot pilot

had devolved. It employs about a dozen people as pilots, but as you said earlier, none of them do it full time. Instead, Jim says, they can use the job as a way to get into the company and then move on to do something else. The current facility for pilots is just a corner of its Toronto office. Our next one will probably be in place like Mexico, and so it might be that there's more of sort of like a labor farm of people that will do that.

And I'm not sure if we if we do, If we stand this up in Mexico as I'm as I'm at least thinking about doing, then um it would there would be less opportunity for them to transition into other roles. But just like Peter, Jim says he's been surprised that the robots are learning faster than they've predicted. The company recently set up a robot at its trade show in Atlanta. When Jim first explained it to me, he didn't even mentioned there was a pilot involved. He just said it

was autonomous. But then I asked him about it and he corrected self, and so there was a pilot there. It was just that the robot was so good at doing everything itself that it barely needed him to intervene. You know, I've been thinking about what I took away from my visits to these robotics labs, and I can't help but come back to this story of the mechanical turk Um if you don't know that one. This is

this famous eighteenth century chess playing machine. It appeared to be automated, but actually there was a person inside actually making the moves by hand. Yeah, that's a great metaphor. I mean, watching Kindred's robots in these factories, you'd probably think that they're doing all this on their own, but it's possible that they're actually being steered by a human in an office very far away. Yeah, and Kindred actually

adds an interesting twist to that story. It starts off just like the mechanical turk with someone controlling it from far way, but then over time it evolves into what the machine was actually pretending to be, and then the chess player who cramped himself inside is out of a job. Exactly. I took one other thing away. I think at first, it seems like there's this really straight line between where we are now and some future where robots are doing everything that you or I can do, or at least

everything we can do physically. But Kindred set that original vision aside, and it's not clear that once the company masters warehouse robots that it will then train robots to do the next thing and the next thing, until we end up with some sort of autonomous humanoids out of a sci fi film. So what you're saying is that what businesses need in the short term or maybe even the long term, isn't these humanoid machines that can do everything we can do, but probably something a lot simpler. Yeah.

I think it's really tempting to think that when we're making robots, were designing them in our own image. But maybe it just never makes sense to get there. I guess it just depends on what we decide to teach them. And that's it for this week's to cooked in. Thanks for listening. Let us know what you thought of today's show. You can email us at the chrystod at Bloomberg dot met or reach out to me on Twitter. I'm at

Joshua brog std and I'm at aki Ito seven. If you haven't already, subscribe to our show wherever you get your podcasts, and leave us a rating in a review. This really does help us reach new listeners. This episode was produced by Pa Gokari, Magnus Hendrickson, and Liz Smith. Francesca Levi is the head of Bloomberg Podcasts. We'll see you next week.