The Fragility of Automation
By Alberto Moel, Vice President Strategy and Partnerships, Veo Robotics
Author’s note: Welcome to the first of what (I hope) is going to be an ongoing and regular series of articles on Veo Robotics’ position in the automation debate. As the capabilities of industrial automation continue to increase, the impact of robotics and machines becomes a topic of broader discussion on employment, inequality, and the future of work. Many smart and thoughtful people have weighed in on the debate, which rages on in trade press, newspaper op-eds, academia, and late-night existential discussions between college students. I hope our additions to the debate increase the signal and not the noise. Of course, your feedback and comments are always welcome!
In April of this year, Tesla was in the news concerning some struggles in getting the highly-automated Model 3 production line up and running. Notably, Musk regretted using so many robots.1
In Tesla’s May 2018 quarterly earnings call, Musk elaborated further, referring to a “flufferbot”:
“We had these fiberglass mats on the top of the battery pack. They’re basically fluff. So we tried to automate the placement and bonding of fluff to the top of the battery pack. Which is ridiculous. So we had this weird flufferbot. Which was really an incredibly difficult machine to make work. Machines are not good at picking up pieces of fluff. Human hands are way better at doing that. So we had a super-complicated machine. Using a vision system to try to put a piece of fluff on a battery pack… The line kept breaking down because Flufferbot would frequently just fail to pick up the fluff. Or put it in a random location.”
Here at Veo, we were unsurprised that the automated production line hadn’t lived up to Musk’s promises. Industrial automation, contrary to much conventional wisdom, is actually quite inflexible, rigid, fragile, and difficult to implement.
We certainly agree that automation is superior to human performance when it comes to speed, reliability, precision, and endurance. It’s also clear that automation can decrease defect rates, reduce material waste, increase reliability, and reduce product variability—all good things from a cost and efficiency perspective.
But automation is expensive; it involves a lot of engineering resources, overhead, and manpower to test and install. And although a well-designed automated process improves manufacturing efficiency, it is fragile and naturally lowers system flexibility. When automating a manufacturing step, everything must be placed precisely, with each part in its proper location. There is very little room for deviation in the parts and the process. Fixturing has to be just so, and any failure to abide by the exacting manufacturing process design will trigger an error, a line stop, and a need for human intervention. In the worst case, as the Tesla Model 3 production line demonstrates, a small bug can lead to a complete meltdown of the manufacturing process.2
One of my favorite examples of this subtle combination of rigidity and fragility is something I witnessed first hand at a robot factory in Japan. The factory ran a highly-automated robot manufacturing operation where robots and automation systems carried out the manufacturing of their robot offspring, and where 80-90% of the assembly steps were fully automated.
I saw up close a robot mounting a cover on a "child" robot arm by bolting it with six screws. To do this, the robot would pick a vertically-mounted screw (screw head on top) from a specially designed screw carrier using a pincer arm (the screw carrier itself was loaded and prepped by a human). It would rotate the screw 180 degrees from the original orientation (screw head now on the bottom) and mount it in another fixture that would turn the screw 90 degrees. The robot would then retract and switch arms from the pincer to a hex screwdriver, pick up the screw from the fixture, and bolt it on the "child" robot.
The entire process would take a couple of minutes, and it would be repeated for every screw on the cover. It was like watching a slow-motion insane mechanical ballet put on by Dr. Evil of "sharks with laser beams" fame.
A capable human could have put on all six screws in the time it took the robot to mount one, and the human could have picked the screws out of a loose bag, without the need for any fixturing, programming, or safety interlocks. The only benefit of automation that I could see was that the robot could apply a very precise torque to each screw, but that could be solved by giving a human the right tools, such as a torque-controlled electric screwdriver.
Further, if one of the screws and bolt-holes weren’t lined up just so, or the screw had a defect or was of the wrong size, pitch, or even the wrong color, a human could very easily recover without missing a beat. But the rigid system used to automate this process would not know what to do (unless it was explicitly programmed with a solution) and everything would melt down and come to a screeching halt. And then a human would have to step in and fix it before restarting the production line.3
Although the automation of the screw-mounting operation led to very low product and process variability (important attributes on their own), it did not simplify the manufacturing process. Far from it. The automation actually introduced dozens of redundant process steps into the workcell, all because of the robot’s natural inflexibility. The additional complications might have lowered process variability and increased product quality, but they also increased system fragility.
Highly automated systems are brittle and have limited room for error. In the language of systems and signals engineering, they are high-frequency systems, where the manufacturing process has to account for many specific and determined steps, with little room for deviation from a specific program or routine. If you were to imagine a “system clock” that would run such a process, this clock has to “tick” much more frequently to account for the incremental process steps introduced through automation. Every time the clock ticks is one more chance for something to go wrong.
As the example above illustrates, humans in the loop can apply their dexterity, judgment, and creativity as dampers to the high-strung, pins-and-needles4 nature of advanced automation. They can serve as smoothing functions in a process that otherwise would not tolerate any ambiguity or deviation. With humans in the loop, we can make the system a low-frequency one, where the tick-tock can be slowed down as humans perform tasks in the process that would have to be over-determined if fully automated.
Automation can substitute for labor and, in many instances, it’s the right approach. However, allowing humans and machines to complement each other is likely to lead to more robust, less brittle manufacturing processes.5 Developing manufacturing processes that incorporate a human in the loop from the beginning are likely to be less prone to tantrums and meltdowns, as humans can “correct” for small process variations on the fly.
But what about the economics? Even with brittle highly-automated systems, the economics of automation may favor more machines and fewer humans...or maybe not? That is the subject of our next couple of articles, so stay tuned.
2 There is even a book dedicated to the failure of overly complex systems, Meltdown: Why Our Systems Fail and What We can Do About It, by my friend Chris Clearfield and András Tilcsik. Chris, consider this a plug for your book.
3 A retort to this inflexibility is that it’s actually a temporary condition and technology is evolving rather rapidly to give robots additional judgment and flexibility to be able to pick up objects and understand their environment. But all of this is still bounded by human judgment and what we program the robot to do. For example, at the Japanese robot factory, we could incorporate exception handling and “intelligence” to allow the robot to understand screw shapes and thread counts, but if we give it a screw of a different color or material (e.g., aluminum instead of forged steel) and we do not program that as an exception handling routine, it would never know this. A human would immediately know something is wrong and discard the part. The autonomy required for a robot to understand color or density isn’t going to happen unless humans give it to the robots. David Mindell at MIT has taken this argument into a book-length disquisition (Our Robots, Ourselves: Robotics and the Myths of Autonomy). More broadly, no “Singularity” is going to give a machine the ability to understand and act on its environment outside of what it’s been “programmed” to understand, in the same way humans can’t grow a third eye or another limb at will. It’s not encoded as a capability in either the machine or the human.
4 Literally, not just figuratively. Advanced automation systems usually include very precise fixtures and jigs incorporating multiple positioning pins and fixturing.
5 This is what Mindell calls the myth of replacement. In his words:
“the idea that machines take over human jobs… Human-factors researchers and cognitive scientists find that rarely does automation simply “mechanize” a human task; rather it tends to make the task more complex, often increasing the workload (or shifting it around).”