Working With Robots in a Post-Pandemic World
29:02 Share

Hello and welcome to SkyNet Day's Let's Talk AI podcast, where you can hear from AI researchers about what's actually going on with AI and what is just clickbait headlines. We release weekly AI news coverage and also occasional interviews such as today. I am Andrey Krennikov, a third-year PhD student at the Stanford Vision Learning Lab and the host of this episode.

In this interview episode, we'll get to hear from one of the authors of a recent article, Working with Robots in a Post-Pandemic World, Professor Matt Bean. Professor Matt Bean does field research on work involving robots to help us understand the implications of intelligent machines for the broader world of work.

Bean is an assistant professor in the technology management program at the University of California, Santa Barbara, and a digital fellow with Stanford's Digital Economy Lab. Thank you so much for joining us for this episode, Professor Bean. Happy to join you. Alrighty. So, yeah, our focus here will be your article, Working Robots in a Post-Pandemic World, which just came out a few weeks ago.

Before we dive into any details about it, maybe I can just let you go ahead and provide a quick summary of sort of overall what the article is about and what are its high-level conclusions. Sure. So I'll start with the high-level conclusion, actually, which is that firms now that are in our sample, at least, and this finding seems to generalize as I've talked with others across other industries,

are facing a lot more uncertainty and volatility than they're used to. So our particular research program is covering e-commerce, for example, you know, fulfillment warehouses, but also package or parcel transport. And the entire world on many dimensions is really spinning. And so traditional ways of appraising investments in automation are really

but just not applicable currently. When firms are looking to make investments, the traditional calculus for what makes for most value per dollar, both immediately and in the long run, are thrown off kilter mostly because of change. Volatility is basically the enemy of automation, really elaborate or complicated automation. So practically, firms are

Turning more is, of course, varies by firm, it varies by industry, but turning more towards automation that we coined this term, we classify as plug and play. It has a certain number, a sort of set of characteristics that

But the basic ones are it's rapidly, you can sort of buy it and have it delivered rapidly and it can be repurposed rapidly. It's reconfigurable. It's modular and interoperable, highly modular and interoperable with currently available automation. And it doesn't require a lot of training time.

Another interesting piece, at least for us, is that it has a small physical footprint. So there are some automation that might meet all those other characteristics, but square footage in a facility that automates some sort of physical process is quite the

scarce commodity, and also some changes with respect to automation footprint have implications for other adjacent processes that have automation involved in them.

And, uh, changes are not simple. Changes involve bolting things into concrete or shearing bolts out of concrete, uh, moving power, air supplies, uh, redirecting, uh, conduits and chutes, that kind of thing. Uh, all of which takes a fair amount of calories, time, uh, and dollars and so on.

So it's given the dynamism in the market and so on right now, it's just become much easier and more appropriate for businesses to invest in this kind of plug and play automation that if

skew changes next week, if a work process changes next week, we can quickly repurpose those investments to serve that changing demand. A classic example I trot out in interviews now is a simple industrial strength pump for

for moving fluids of various viscosity. There's a little bit of software on board. There are some sensors, but certainly no AI, nothing fancy. But it could be used to nail polish, hand sanitizer, soup. You can go on and on. And firms are needing to pivot that radically these days, some firms anyway. And so that pumps in extremely good investments.

Anyway, so that's the sort of core finding. The broad study, obviously, we didn't intend to go studying the implications of a pandemic for automation. We started this study about a year and a half ago, maybe a little bit more.

covering now going on eight firms that are deploying AI-enabled robots for repetitive manual work in warehousing almost exclusively, and then their customers. And we're asking different questions. We're asking questions about warehousing

When and how do firms and then frontline workers adapt particularly constructively to some discontinuous, qualitatively new automating technology? It just so happened that we were in touch with these firms and these deployment sites for a good year and a half, and then COVID came along. So we couldn't go visit those sites. We couldn't call them. We couldn't ask for more data without asking about the implications of COVID. It would be...

it would be really inconsiderate not to. I mean, these are people we've come to know and like and trust and vice versa. So we just started to collect a lot of data along those lines and noticed consistent trends across firms of various sizes. And we've covered almost all geographies in the U.S., different industries, and the patterns were the same. So that was the impetus for writing up this piece for Sloan Management Review. I

I see. Yeah, that makes a lot of sense and is very interesting. I think part of the motivation in the piece starts out by noting that in other media articles, such as the New York Times article, Robots Welcome to Take Over as Pandemic Accelerates Automation, kind of the intuitive story has been that

Because people need to work remotely and because of the pandemic and different things, there will be an increase in automation, right? And more robotic adaptation and maybe even more AI driven robotics, wherever, you know, really advanced systems. So it sounds like your story here is interesting.

that it's not necessarily the case that companies are being able to adapt very sophisticated advanced robotics and replace existing procedures more so it's these specific plug and play type robots is that right yes yes that's true that was always roughly true before covid so um

We know already that, you know, especially from some of Eric Bernielson's recent work with some colleagues and the U.S. Census, that it's, you know, their finding was 1.3% of surveyed 800,000 plants in the U.S. were using robots of any kind.

Now, this is to say nothing about AI-enabled robots. So that's probably 1.3% of that 1.3%. So this is an extremely uncommon technology, and it's underproven, is what I would say.

So, and by under proven, I just mean it's not reliable enough, simple enough to implement at scale in a way that is where the cost, it's cost effective, where it's a good investment for firms yet.

And so it was always somewhat sort of, especially for larger firms, many larger firms make repetitive bets on under-proven tech like this to try it out, to figure out where and how it might be useful. Co-development projects is one term that

in a variety of disciplines is sort of an understood bucket for this kind of activity. Right now, the minute COVID hit, the headwinds for those kinds of projects just became much stronger. They're still going on. We still have deployments to study. We still, you know, amazingly, none of the firms, the vendors in our sample have gone out of business yet or run out of funds. I expect that may happen, though, because, you know,

Experimentation takes time, it takes resources, it takes physical square footage. And you know, for example, if you have a robotic system that is doing pick and place work, you know, sort of taking sample, you know, sort of sample goods out of a bin on differentiated sample goods and putting one of those in a box as it goes by on a conveyor, for example, that's a prototypical example.

It's an amazing achievement to watch. I mean, just technically, it's stunning that these systems can do this with any reliability, anything approximating sort of industrial grade reliability. But setting that up requires stopping a traditional production line. So you have to take humans off of that line, set it up as a robotic experiment, essentially, which means, let's say, if you have 10 lines and you take one down, you've basically

essentially depress 10% of your product, you know, your production lines or production capability, at least for the time of that experiment, which probably will be say three to six months right now, if your demand has just quadrupled or octupled, uh, you, your op, your operations people will never allow that. They will, um, their job is to make sure that they deliver for the customer and taking down 10% or even 20 of your capacity, uh,

to try out, these are not my words, there's some sort of science experiment. It's often how they're referred to. It's just much, much harder to justify right now. Yeah, it makes a ton of sense. So with AI and robotics, of course, often the story or

people have concerns about job loss, that robots and AI will replace human workers. And, you know, within a decade or even less, you know, a lot of automation will come for human labor and jobs. So I wonder, yeah, if you can expand on the implications of your findings before COVID, during COVID for robotics and human jobs. Okay.

Sure. So a few caveats and then I can share some points. So the main thing to say is that we'll have more to say on this question, the broader sort of implications of AI-enabled robots for work and employment and so on in subsequent papers. So stay tuned is the main message. I can start to speak at Ground Zero, which is COVID, and this paper that is published right now.

The clear implication, the sort of the concomitant finding with what I've said already around plug and play automation is that human flexibility is more valuable now than it was before COVID. There's a lot made out of trying to create more social distance within these facilities. And that is important and it is hard and it does

Make it on some levels attractive to try to automate people out of a process to avoid contact, right, to avoid transmission of the virus. Practically, that is more achieved with plexiglass and cardboard than it is with any new kinds of automation or just extending the length of conveyors just to separate people a bit more.

But those kind of effects are dwarfed by the need to cope with change and uncertainty in the work. And the way you do that is to, instead of try to automate in a building, is just fill that building with people. So I have a number of organizations in our study. Now, these are consumers of robotics, sort of mid to large scale organizations.

fulfillment type organizations that have stood up greenfield sites, you know, new buildings in certain geographies or are repurposing old buildings. And they have scrapped their automation plans entirely. And it just said, look, add 20% more people to that building. Yes, it's more costly. Yes, it's harder to manage. These things are all have costs associated with them.

But we can tell people, people are exceptionally good at coming up with new ways of handling surprise new products, adapting to change, basically. And they are, you can also, if you do a good job managing a team, managing an organization that sort of fills a building,

you can motivate them to work harder to meet some new challenging goal. You know, a lot of us in very different ways are, you know, this is an extraordinary challenge, this pandemic, but we're seeing some of the best of what people can do for each other in their work because of that challenge. And we've seen that in a number of facilities as well. So this is a long way around to say in the short run for people,

and jobs, there's more opportunity, not less. There's more jobs that need to be filled. And within a job, if you or I got a frontline job in one of these facilities, that facility now is investing less than it used to in complicated automation that will sort of wipe out entire work processes, or at least radically reconfigure things such that you don't need as many people.

So you or I not only have more job opportunity at the front line, but we have more opportunity to be valuable, to innovate in a process, to note some way in which it could be improved, to learn more about the functioning of the building, to connect more with other people. So a highly automated process is one where the connections between humans are

are physically distant and socially distant. That's what a highly automated process looks like. You don't get to talk much with each other. You don't get much to socialize or on the sort of learning side, you don't get much to sort of notice things as a collective, like chat about how this might be improved, see new opportunities. And now there's just a lot more of that. So in a weird way, COVID has provided more

more job opportunity, more opportunity for learning and development, but certainly more opportunity for work than there was before.

I see. Yeah. So that's, I guess, kind of important or at least useful to be aware of that, even though intuitively could sort of see how people might lose jobs and be replaced with automation that isn't susceptible to the disease. You show in this study that, in fact, in the short term, during COVID, in response to the pandemic,

That's not the case. And primarily, it seems that these smaller plug and play systems are what people can use and the larger, more advanced, more complicated things are not what you rely on in an emergency situation. And I wonder just to ask you, is the conclusion sort of that broadly when we have these sort of crises, that's what we can take away about automation and use of robotics?

Uh, that's interesting. Um, that's an interesting potential implication that deserves further study. It's, I think that stands to reason though, right? Um, uh, there, there's an entire literature on how, um, organizations, groups, and societies respond to crisis, disaster, you know, things like disaster, um, or massive opportunities, right? So, um, let's go to the moon or there's a gold rush across the United States, um,

And so on. And the consistent finding across that body of work, as I think about it, is that

complexity and rigidity in the way you organize, in the technologies that you use is kind of the enemy because there's a lot that you need to learn really fast and you need to be able to adapt, i.e. change process, change personnel, change skills, given rapid change. And so, yes, I think that's, you know, our findings are a variant of that.

just a bit more fine-grained set of findings around the implications for automating technologies. But as long as that volatility is there, I think we're going to see a lot more stories about, and a lot more data about humans playing a really important role around, you know, adapting and finding new ways to handle things. Yeah.

I see. Yeah. So as usual, I don't underestimate humans and how important they are still, even as we do make various advances. There's a quick important thing to add on there, though, which is that I think we are all, you know, the intuition that AI enabled robotics will at some point in the relative mid or even longer term, somewhat longer term,

be able to do repetitive manual work in highly uncertain and dynamic conditions at high reliability. Having seen what I've seen, I can't speak in detail about these firms, but having seen their deployments, I've been absolutely, I mean, I follow these kinds of things as closely as I can. A number of them are doing things that say five years ago would have been written off as not worth pursuing.

because it's just too challenging. So the progress, you know, under the hood, close to the ground is pretty stunning. When you show an example to someone, you know, look, this robot can handle these perfume samples with 99.96% reliability. You know, when you change the lighting, when you change the orientation, when you change things like temperature and so on.

That doesn't look amazing to a layperson, but to me it is amazing. And so I think it's a bad idea to underestimate these effects. It's just that they're not here yet. It's going to be quite some time before these things are cost-effective at scale. Yes, that's a good point as well. And I actually want to touch on a related piece of work cited in the piece from Eric. Eric, I actually don't know about your series name.

From Erik Brynjolfsson, the paper, Reproductivity J-curve, how intangibles complement general purpose technologies. So, yeah, I also wanted to quickly touch on it sounds like a general kind of story is that even as you get these new technologies like these AI enabled robots,

Even when you have a proof of concept and you have initial results, it'll still take a large amount of investment and a fair amount of time for the firms to be able to use them effectively and to really get the most out of them. So at first, it'll take years to sort out how to best use them, and then there'll be kind of a big change. Is that a fair, broad kind of guesstimation of how we might see things?

It is both, you know, all models are wrong. Some are useful. That is a useful model that is deeply flawed in one way. So the utility is there. We have studies going back, say, more than 100 years, but certainly 90 or so, showing that any time there's some qualitatively new form of automation that rolls along, it's

It takes, you know, there's no sort of firm timescale, but it takes far longer for it to find its ultimate or a very high sort of maximum of utility for firms and so on. Much longer than anyone would expect. A lot more failure, sometimes death, certainly injury, a lot more money involved in finding those maxima.

And yes, it's because everyone's trying to figure out where and how to use it. The classic example is the electric dynamo. It rolled around around the turn of the century, business started to try to incorporate it, but it took almost 25 years for firms to go from installing a single giant electric dynamo on the single camshaft in the middle of a factory that drove many, many belts that each went out to a separate piece of equipment,

to attaching a small dynamo to each piece of equipment. So yes, that's a sort of very generalizable finding. The key falsehood in there, and this is where my work kind of compliments Eric, is that it is necessarily true that someone figured that out with each of these new technologies far earlier than anybody else, either in whole or in part.

So it's logically impossible that everyone simultaneously 15 years later discovered this single dynamo to single piece of equipment solution. Someone did it earlier.

And it just took a while for that to diffuse. There are many reasons for that. But the kind of work I do is to go out into the field. My prior study with robotic surgery was the same, basically. Even right now, a good 15, 18 years after its introduction, trainees are struggling to learn how to use the robotic surgical system known as the DaVinci.

And other robotic surgical systems as well. But a very few found new ways to train and learn how to use the system. It's just they're quite rare. They're isolated from one another. They don't recognize that they're doing anything innovative. And so it's hard work to find them. That's the goal with our current study is to go out and to include a large number of companies, vendors in our sample and an even larger number of deployment sites to try to find

individuals, frontline workers and organizational practices or conditions that lead to discontinuous early success compared to everybody else. No one organization or individual is going to get everything right and do it all perfectly in the way that we would all recognize as success 15 years later. But, you know, if you have a big enough sample size, you can find that positive needle in the negative haystack.

And so that's the way in which this sort of J curve finding is accurate, but also obscures that, you know, success doesn't come from nowhere all at once. Makes sense. And sounds very exciting to hear of your future research and what comes out there as far as seeing the early adopters who really make the best of what we have today, which, as you know, are some really impressive advancement in AI enabled robotics.

So we have covered a good deal, I think, of what has been an article and some of the related topics there. I guess I just have to ask you, is there anything we haven't touched on that you think is worth noting with respect to robot use post-pandemic?

I think it's, you know, I suppose the thing to say is the reason I got excited about this kind of study, I think the reason Eric is, I think the reason the vendors in our study and the organizations in our study are excited is that

There are a lot of claims out there about the implications of these kinds of technologies, particularly the very advanced ones for work, employment and so on. The sort of the kind of society that we end up living in. And we have a lot of questions about are these technologies being built, sold, deployed, consumed, used, modified in ways that are

move us closer to a society that we would like or farther away or both in some cases. And we just don't have a lot of good data about that. I mean, we're starting to, we're starting to get some interesting studies, but I think it's especially when these kinds of

have strong political implications where big decisions about funding, education, government, law, and so on are being made and have been made for the last decade absent this kind of data. I think, at least for me, this is the real motivation behind the work. And if there's one message I would hope to spread, it's that

We should all try to get more data on what's actually going on in the ground, whether it's to do with design processes, processes through which these technologies are sold to the market, to the open world, the processes through which they're put to use, and so on. Getting data on these processes will help us make actual informed, do-good informed processes.

social science, like build good theories about what's actually happening out there. But then also for politicians, for lawmakers, for business leaders, for technologists to make

informed decisions about what the consequences will be of these technologies, I think we need a lot more of that kind of work. Obviously, we all say, if you've got a hammer, everything looks like a nail. And this kind of work and getting this kind of data is why I do the work that I do. So, of course, I'm going to say something like this, but it does seem to me to be a serious gap

And making decisions without this information, all you have to go on are your assumptions or an anecdote that you've heard about what's going on, not a systematic data set. So I hope if there's anybody out there listening who feels the same way I do, get involved, try to support a study. If you want to be studied, reach out.

You know, we just need to build a bigger tribe of people focusing on this kind of work so that we can make decisions that are more fact based and data based than based on mental models that, you know, they're based on a variety of things. Science fiction, we've seen stories we tell in here, you know, you name it.

Absolutely. And as an AI researcher and someone focused on robotics, where I see these stories all the time and, you know, people mention them, I'm very glad to see that you are doing this on the ground study and looking at the actual implications and collecting that. So on that note, we can go ahead and finish up. Thank you again, Professor Bean, for joining us on this episode. Happy to join you. Really appreciated the opportunity.

And thank you listeners for listening to our interview for this episode of Skynet Today's Let's Talk AI podcast. You can find articles on similar topics to today's and subscribe to our weekly newsletter or similar ones at skynettoday.com. Subscribe to us wherever you get your podcasts and don't forget to leave us a rating if you like this show. Be sure to tune in to our future episodes.