The Hidden Costs of Full Automation

By Alberto Moel, Vice President Strategy and Partnerships, Veo Robotics

In my previous post, I made the case¹ that full-on automation is inflexible and fragile and leads to meltdowns and all kinds of flops and misfortunes. One would assume that these “line down” conditions would cost money and lower the financial returns to automation. But what about the upside? Would the higher throughput, precision, repeatability, and speed make up for these line fails so that, all-in, it’s economically sensible to have high levels of automation? In other words, even with the risks and inflexibility of automation, is it the right economic choice?

It turns out that, in general, the answer is no. Is it then better to eschew automation altogether and just go old-school, using humans for everything? The answer, again, is no. It turns out that there is (in theory and in practice) an optimal combination of humans and machines working together that’s the most productive.² To make the argument, let’s ease into a couple of qualitative examples before reviewing more general data points to clinch the deal.

Let’s start with something many of you have frustratingly experienced: a broken windshield on your car. Nowadays, a man with a van³ shows up at your home or workplace, takes the broken windshield out and replaces it with an aftermarket one. Notice that I said a “man with a van,” not a robot in an automated guided vehicle. But if you were to visit the factory where your car was made and stop by the workcell that puts the windshields on those very same cars, it is very likely that you’d see a robot or other forms of automation mounting the windshields and repeating the process over and over. What’s different between these two conceptually similar processes?

In one case, the car manufacturer is putting out a car a minute, so spending hundreds of thousands or perhaps millions of dollars to automate the step makes sense: the direct and indirect costs of automation are well amortized over the units of production over a run possibly in the hundreds of thousands of vehicles. Just as importantly, putting on the windshield for the first time is an extremely well-defined process. The windshield-less car comes down the assembly line in the exact same position and state as the next car down the line, everything is nice and scrubbed clean, there are all kinds of fixtures and measurement sensors overseeing the proceedings, and the process steps are consistent and well specified.

Now consider the case of the unfortunate rock-paper-scissors game the windshield of your souped-up BMW just played (and lost). There’s probably broken glass all over, which needs to be carefully vacuumed; the seal around the windshield is likely torn or scratched; there may be structural damage to the car that needs to be inspected; the work will probably need to be done outdoors in your driveway or workplace; and the right parts, fittings, and tools for your very specific car model (not one of thousands, but specifically yours) must be prepared. It doesn’t take a genius to conclude that assessing the situation and making the windshield replacement is going to require far more adaptability, judgment, and manual dexterity than the assembly line windshield install. And I can tell you, with total certainty, that there is no robot out there with the “smarts” to do this, at any cost.⁴

Clearly, even for the same step, depending on the environment, complexity, and repeatability of the activity, sometimes it’s best to use humans, and sometimes it’s best to go all-in with the machine. That debunks the idea of automation always being the best outcome, but how about an optimal combination of humans and machines? To apply some intuition there, let’s go to another example in automotive, one that is usually done in full automation, but requires a bit of human love in the process: car body welding.

Car body welding is one of those manufacturing activities that robots are far better at than any human ever will be: robots are quicker, more precise, and tireless, and they don’t mind the heat, high voltages, and general nastiness of the process. Many of you have seen the canonical “turkey farm” welding station⁵ with robots working silently and without complaint, nary a human in sight. But after a number of cycles (as often as a few hours or maybe a couple of days of production) the welding tips need replacement or cleaning as they naturally wear out.

One approach is to build a set of additional fixtures and automation just to clean and replace the tips. But this costs a lot of money and the actual use of these specialized components is limited to at most a few minutes a day. The automation doesn’t pay, and it’s cheaper to shut down the line periodically (at $50,000 a minute) and have a human step into the space to do the dirty work. Although this step involves a complete line shutdown, which is itself quite expensive, it’s still more economical than automating. This approach would be even better if humans could enter the welding workcell while the line was still active (not directly around where the human is, of course, but elsewhere up and down the line). Then the cleaning and replacement could be done much faster.⁶ So a combination of machines and humans working together would be superior, in this case, to full automation.

These two examples show that automation can be enhanced by human collaboration—the simultaneous use of the strength and speed of machines and the judgment and dexterity of humans is the most productive. Other data points bear this out, and this is no secret to any seasoned manufacturing engineer who has been dealing with these human-machine interactions for many years.⁷ Take, for example, Gorlach and Wessel’s examination of the optimal level of automation in the automotive industry.⁸ Their model and empirical observations (applied to VW plants in Wolfsburg, Germany, and Uitenhage, South Africa) tell us that, beyond a certain level of automation, manufacturing costs begin to rise, and there is an “optimum” level of automation for a given manufacturing process (Figure 1).

To put it in simpler words, too many humans are too slow and make too many errors, and too many machines are inflexible, fragile, and brittle, resulting in high development and integration costs and more frequent line stops and rework. There is a best combination of humans and machines that minimizes manufacturing costs. This combination will vary by application, of course, but in general, adding humans to an automated activity (or automation to a human activity) will improve productivity.

At the micro level, Julie Shah at MIT has been, for several years, developing ways for robots to interact intelligently with humans. Her research group’s work shows that humans and robots working together are more productive that human-only or robot-only teams. In her PhD thesis⁹ she develops Chaski, a “human aware” robot execution system that makes human-robot interactions more fluid. She runs experiments in which a person works with a Chaski-enabled Nexi, a dexterous mobile robot, to collaboratively assemble simple structures using building blocks. She finds that using Chaski reduces the human’s idle time by 85% relative to all-human teams performing the same task. Our own work at Veo¹⁰ finds about a 2x improvement in cycle time for a simple assembly task using human-robot collaboration relative to using humans only. And more interestingly, recent research from Julie’s lab suggests that people collaborating with manufacturing robots don’t mind when the robots take the lead and guide human actions.

At the macro, larger dataset level, work done in the early 2000s by Frits K. Pil and John Paul MacDuffie as part of Phase Three of the International Motor Vehicle Project¹¹ showed, in a roundabout way, that automation doesn’t necessarily mean higher productivity. This conclusion is more explicitly supported by Ron Harbour, a manufacturing expert at Oliver Wyman (and author of the highly regarded Harbour Report), who was quoted in 2017:

When it comes to what functions to automate, we’ve seen differing philosophies. The manufacturers we saw as most competitive—the Japanese, the Koreans—needed a business case for automation. Whereas automakers in Western Europe and even the United States went more aggressively towards automating things, even when that didn’t pay off. We’ve seen examples of companies installing automation that required more people, with higher skills, than were required before. They did it just to display their technological prowess. I’ve always been dismayed by that. Ironically, the most [heavily] automated factories in the Harbour Report are not in the top quartile [in the productivity ranking]. Many are in the bottom.

Thankfully, Max Warburton and Toni Sacconaghi of Bernstein Research¹² have used IMVP data and Ron Harbour’s quantitative analysis to illustrate an inverse correlation between automation levels and labor productivity. Figure 2 (adapted from Max and Toni’s report) shows hours per vehicle data for specific manufacturing plants by German, French, US, Korean, and Japanese car makers in the EU and US. The evidence is clear: plants that are more automated have lower productivity. Of course, even the most backward auto plant will have a certain level of automation (e.g. painting and welding), but it’s straightforward to conclude that over-automation is counterproductive.

In my next post I will delve a bit deeper and illustrate with a detailed example what kinds of productivity and cost benefits can be obtained from careful human-machine interaction relative to full automation or no automation, so stay tuned.

---

¹ Categorically and with impeccable logic, I might add (but I’ll take it if you were just mildly entertained enough to want to read this post also).

² And you can take that to the bank, literally. You could stop reading now but then you’d miss out on a combination of data points and convoluted examples that will either make your eyes glaze over or fill you with the joy of insight. Up to you, red pill or blue pill?

³ I said “man with a van” because I like the alliteration. I could have very well said “woman with a van,” but it doesn’t have the same rolled-r’s aspect to it.

⁴ And there probably will not be for a long, long time. Recently, researchers at Nanyang Technological University in Singapore put together a two-robot system to assemble an IKEA chair—not any IKEA chair, but a very specific one. It probably took a team of PhDs months to develop and it was such a big deal that it was published in Science magazine. Now generalize that to an arbitrary windshield on the car in your driveway. How long do you think it’ll take scientists and engineers to figure that one out?

⁵ Here is one of the BMW Spartanburg welding line. To me, these turkey robots look more like pigeons eating noodles. Note that the robots in the first video are dumber than the 4-neuroned flying rats in the second video.

⁶ This would be an excellent application of the system Veo Robotics is developing, if you don’t mind the plug.

⁷ Except maybe to GM in the 1980s or to Tesla in the 2010s.

⁸ In a paper coincidentally titled, Optimal Level of Automation in the Automotive Industry (Engineering Letters 16:1, 2008).

⁹ Fluid Coordination of Human-Robot Teams, MIT PhD thesis, February 2011.

¹⁰ The details of which I will report in a subsequent blog post, this one is already long and data-heavy as it is, and I don’t want our readers’ heads to explode, or worse, have them stop reading our musings.

¹¹ Now morphed into the Program on Vehicle and Mobility Innovation.

¹² In the interest of full disclosure, I used to work at Bernstein Research and I am “borrowing” one of their charts from their March 28, 2018 report, Tesla: Model 3 and the fallacy of automation—what we believe is wrong and why it may remain difficult to ramp production. The report is only available to Bernstein clients, and in it they basically hammer the point I’m making in this short blog post that automation in capital goods manufacturing has negative returns to scale.