Technology Executive Council
To join the CNBC Technology Executive Council, go to cnbccouncils.com/tec
Technology Executive Council

Why Amazon warehouses and Tesla auto plants will not go 100% robot any time soon

Key Points
  • Amazon uses automation well, but even it has failed to figure out way to make robots pick goods as well as humans can.
  • Elon Musk conceded that he pushed Tesla automation too far, too soon, and had to pull back on the effort.
  • Recent research shows that only 1.3% of firms have adopted robotics, and incremental investment in robots will be wiser than big bets, in spite of arguments that Covid-19 should lead to a rush to replace human workers in warehouses and plants.
Kuka robots work on Tesla Model X in the Tesla factory in Fremont, California, on Thursday, July 26, 2018.
Mason Trinca | The Washington Post | Getty Images

"Excessive automation at Tesla was a mistake," Tesla CEO Elon Musk tweeted back in 2018 amid electric vehicle manufacturing struggles. "Humans are underrated."

Amazon, for all it advances in automation with warehouse machines like Kiva robots, still can't find a robot that can pick a product with as much success as a human. 

Fears about the era of automation displacing human labor can seem old already, but the failures of automation are nothing new. Long before Tesla's attempt to "hyper-automate" Model 3 production at its Fremont, California, auto plant revealed how robots still struggle to deal with inconsistencies in assembly tasks — and that human worker flexibility remains a critical step in manufacturing — GM failed spectacularly with robotic manufacturing as far back as the 1980s. Four decades later, top researchers say most firms that bet big on automation will continue to be burned. That is in spite of the Covid-19 pandemic adding another argument, at least in theory, in favor of more automation: i.e. robots in the workplace can't contract a virus.

VIDEO8:3008:30
Robotics ROI: Attainable automation in a pandemic and beyond

In an article written for the MIT Sloan Management Review, "Working with Robots in a Post-Pandemic World," published on Wednesday, Erik Brynjolfsson, director of the Stanford Digital Economy Lab, and Matt Beane, assistant professor in technology management at the University of California, Santa Barbara, and a digital fellow at the Digital Economy Lab, write that most companies which investment in automation on a grand scale are likely to see their efforts fail, while those that succeed will be atypical. And while those rare successes should be studied closely, they argue that it is incremental investments — plug-and-play technology models for settings like warehouses — that remain wiser today. It is a conclusion they have reached that does not change based on the size of a company's balance sheet, or as a result of the coronavirus. While the march of automation will not reverse, it has a long, long way to go: only 1.3% of firms have adopted robotics to date.

Brynjolfsson and Beane recently spoke with CNBC about their findings and ongoing research project on automation and the economy.  The interview has been edited for clarity and length.

CNBC: Were Tesla's stumbles with automation a prime example of the argument your research is trying to make? 

Brynjolfsson: It's hard to get it right, what level of automation to dive into. Elon was very ambitious. Most companies are being much more cautious. We did a survey of 850,000 plants [with partners including the U.S. Census Bureau] and found very little adoption of robotics, only 1% of plants are using them so far. Often things don't work out. There are so many processes that need to go along with robots. Technology by itself can't transform companies.

Beane: Robots are a relatively advanced, relatively under-proven technology. Robots are a nice case study on how organizations struggle, and why they under-invest: because they have learned their lessons betting the farm on any one technology. Given the uncertainty of the Covid environment, some firms in our study have set up entirely new buildings and deferred any investment in new technology other than invert conveyors, for example. They just filled these greenfield sites with people because installing expensive automation requires a clear sense of demand in the mid-term. No one can predict that well right now.

Experimentation at Tesla or Amazon, sometimes leads to big, costly mistakes and sometimes it is wildly successful. It is hard to know, sitting in an office, what will work the way you want, and it is not unreasonable for companies to test and push concepts. ... but the tsunami of change is still very much in the future.
Erik Brynjolfsson
director of Stanford's Digital Economy Lab

CNBC: Amazon has stressed that it adds more human workers than robots. Does Amazon, and your research, teach us that even the biggest companies will fail if they bet too big on automation?

Brynjolfsson: Amazon is constantly testing the frontiers of what machines can do. Humans and machines have different strengths and weaknesses, and it is always better to have some combination of humans and machines.

Amazon's distribution system has some very sophisticated robots that move products inside the distribution centers, but for pick and pack tasks, they have not found a sufficiently dexterous robot to recreate what a human can do. They've also found that they can't scale up and down with robots as rapidly as they can add more people at times of demand spikes, like holidays. Robots have basically one speed and they are hard to ramp up during peak periods. Amazon is constantly innovating and working on better technology for manipulation and I am sure over time we will see more tasks done by robots. But they are being smart: regardless of the amount of money you have, you don't want to waste it if working with humans gives the company a better return.

Beane: Large firms with lots of capital investing for their long-term sustainability and innovation will remain relevant. It would be naive to think otherwise. They have to try something new, and it wouldn't at all surprise me if firms that are making risky bets now did better in the long run — say five years from now — as a result. But for most firms in our study, in the midst of a pandemic, the headwind to experimentation is stronger.

Now is a time to get immediate, more proven results from plug-and-play technology — think, for example, of a machine that folds cardboard boxes or pumps that can be repurposed to handle production of nail polish or hand sanitizer, useful for many jobs. There are a few examples of robotic technology that are extraordinarily reliable and easy to implement such as some materials transport robotics. You can buy those robots all day long and know exactly how to implement and they are reliable, and cost-effective.

An Amazon Robotics robot moves a rack of merchandise at an Amazon fulfillment center on January 20, 2015 in Tracy, California.
Justin Sullivan | Getty Images

Amazon's Kiva is like this — not experimental. They can add 50 more Kivas and that is a plug-and-play activity. What makes it plug-and-play technology is that it is small, repurposeable, modular and interoperable, and you can get a return from it running right away, and change its use case next week. If next week, it's not sanitizer but jam or nail polish, you don't have to reconfigure an entire line. You don't have to shave steel bolts out of the floor.

But an entirely automated storage and retrieval system moving massive amounts of products around, costing $10 million to install and that can take a year to get going is harder to justify right now. And robots picking items out of a box ... those are under-proven and require lots of effort and experimentation. If you are looking for a good place to get a representative, diverse sample of how an entire industry is dealing with a new class of technology, the whole distribution, packaging and shipping sector is an excellent case.

CNBC: It sounds like you could be talking about co-bots, an idea that companies like as a feel-good narrative about man working with machine. Is that not just rhetoric, corporate PR, but the reality? 

Beane: I nearly did my dissertation on [now shuttered] Rethink Robotics' co-bot Baxter. There really is no accepted definition of what a cobot is, and, in fact, no widely accepted definition of robot or even AI. Typically, these are systems that are force-compliant: you can safely be in a space with them and they won't injure you. They have some mechanisms to operate safely on tasks with humans in the same space. But they are not deployed at scale for businesses yet.

Brynjolfsson: There is a whole spectrum of having humans and machines work together. You can choose different points on the range in this division of labor, but in almost every case, you want humans doing some of the task. Some very repetitive tasks you are likely to automate, but there is a long tail of idiosyncratic tasks that humans can more easily handle. Trying to have a machine learn how to do an infrequent task is just not cost-effective. It is unreasonably expensive now.

If robots are designed to work alongside humans, one way is to have them segregated, for instance with a wall or fence that separates humans from the machine. But before they can safely work side-by-side, machines need to be almost intuitive about where the humans are and need to be force-compliant so they don't injure them.

CNBC: You write in your MIT Sloan Management Review article that the current supply is short of demand for plug-and-play technology, and vendors are likely to cater to existing customers rather than take on new ones. So what can firms do to overcome that challenge?

Beane: Supply is a significant impediment on many fronts, especially now. We've seen firms switch from plexiglass barriers to cardboard, for example, because even plexiglass is in short supply. And that's an extreme example ... under supply constraint, the general move seems to be to less complex, more repurposeable technology. More complex technology is harder to scale, so vendors are struggling to meet demand. It is hard to say how long it's going to take for this imbalance to get cleared up, as these issues cascade through the supply chain. For example, some plug-and-play firms are struggling because social distancing has slowed their steel suppliers.

CNBC: Where are robots getting better at tasks today?

Beane: Part of the value proposition in this investment is doing the same task, say packing makeup samples into a box, 500 or 1,000 times a day, but with a bit of unpredictability. The samples could be slightly different, or the bar code might not be showing, or the lighting might change. So it is repetitive but variation in the conditions meant that most robots, until two to three years ago, could not handle it. But now that's changing and that's a place where the vendors in our study are commercializing technology and having some success.

But systems like these are still under-proven. There is no firm that has solved things completely. Not even Amazon can handle pick and pack work reliably and with flexibility. It is a small fraction of companies even trying anything as forward-looking as systems that could handle the makeup sample example.

Brynjolfsson: Amazon tried and tore it out, because they realized it wasn't working to their standards. That's happens sometimes and that's OK. That's the nature of progress. Experimentation at Tesla or Amazon, sometimes leads to big, costly mistakes and sometimes it is wildly successful. It is hard to know, sitting in an office, what will work the way you want, and it is not unreasonable for companies to test and push concepts.

But we still don't have machines with the dexterity and flexibility humans have to do the full range of tasks. The idea we are in a robotics revolution is maybe true for a handful of companies around Silicon Valley, and elsewhere, but as the Census survey showed, the tsunami of change is still very much in the future.

People underestimate how important it is for successful automation to have long production runs without lots of variation, and a controlled environment. People think a more powerful machine can do lots of things, but more than half of the answer is rearranging the environment. People can figure out how to overcome bad lighting and crooked components, but if you control those factors, you make it much easier for machines. The fastest race car is useless if you just plunk it down in the middle of a rainforest. It needs smooth highways and an infrastructure of gas stations, not to mention a trained driver, before it can reach 200 mph.

CNBC: How about the the food service industry? We've heard in the past about the automation of the "french fry guy" and the front of the house, as it is called, is now adopting the order kiosks. 

Brynjolfsson: If you're willing to have a standardized production line then it gets easier, but if you want to deal with a variety of ingredients and fresh ingredients, the costs can become prohibitive. It's not a matter of buying some new robots and popping the things into the same old way that the rest of the organization makes money. Complementary investments, including new equipment, human skills, product mix and even marketing approach, will be required and could be 10x that of the cost of the robot itself.

VIDEO4:2604:26
White Castle CEO Lisa Ingram on reopening and automation

Beane: As far as deployment in food service, there is no firm betting the farm on tech like this because it is unproven, but five years ago, to imagine a robot on a food service line doing some aspect of food prep was inconceivable. The fact that firms are even experimenting with this, with some reliability at all, is pretty amazing given the unstructured environment and target of manipulation [soft food items] about as unstructured as you can get.

CNBC: How does a firm know where to draw the line between success and failure?

Brynjolfsson: One classic goes all way back to GM in the 1980s when they had an earlier wave of technology and sought to aggressively automate factories. It was a catastrophic overshoot. We have seen that over and over again, but companies also undershoot when it comes to technology investment. Walmart and Amazon advanced because they used sophisticated information systems in the 1990s while their competitors hesitated. You can make mistakes in both directions. That's a pattern we see over and over.

Beane: There are many ways to have success in the short-run with plug-and-play technology. But that might have nasty second order effects. One obvious one is a bunch of plug-and-play solutions might limit the capability for an organization to handle an external curveball. You can deploy automation that gets people and technology more specialized in a way of working that leaves them less cross-trained, less able to pivot to handle new customers and product lines. Some organizations in our study are paying lots of attention to this and rotating people and techniques in tasks so they are better able to handle the next surprise. They are explicitly aware they can't just have success by surviving today — they have to survive in a way that enables long-term success. That's much harder.

A recent survey of 850,000 firms published by the National Bureau of Economic Research found low use of robotics in the business world.