To Be Automated: Finishing

Buffing Wheel End-Effector

Robotic Polishing Setup developed for A. Zahner Company

In the final edition of To be Automated, I recount my summer job in Kansas City working with union metalworkers to automate the task of polishing.

Robotically Polished Stainless Steel

Making Mirrors

Summer 2019

For my first summer job in grad school, I wanted to take a break from architecture. The opportunity came up to work with the Research and Development team at Zahner, an architectural metalworking company. Zahner has their own six axis robot, a de-commissioned red industrial robot from Tesla. The robot arm is the same model of KUKA robot that I learned to use in undergrad, so I could get started right away. Working under James Coleman, Zahner’s lead R&D engineer, I was to spend the summer developing scripts and hardware to equip their robot arm to take over the job of polishing. Specifically, mirror polishing. Everybody loves mirrors, except the people who have to make them.

Division of Labor

The most interesting part of working in the context of Zahner’s manufacturing facility was the division of labor. Each facility, whether it was dedicated to assembly, production, or art restoration, was divided into two parts. On one side, the office was where the digital work happened, and then the shop was where the physical work happened. There was surprisingly little overlap between the people in these two groups, largely due to the union influence in Kansas City. To operate in Kansas City, Zahner needs to work with the local union of metalworkers. The union aims to ensure consistent work for their members, and so they get first priority on any jobs associated with the physical act of metalworking. The union metalworkers are masters at their craft, but most of them are not trained in digital fabrication methods. The workflow for a Zahner project involves the office-workers who are mainly engineers and some designers creating robust digital models and fabrication instructions. These are then handed over to the union workers who handle all of the production and assembly. Numerous perplexing scenarios emerge from this situation. Firstly, the union members almost never know the full scope of the tasks they are performing, because they are not involved in the planning of the digital models. They execute the instructions given to them with near perfection every time, but they don’t have complete information and aren’t going to be able to predict issues down the line of assembly. Secondly, the office workers are not permitted to do any sort of physical labor to maintain union relations, so they design complex assemblies in their computers and then they are executed like magic by metalworking magicians. Furthermore, if the office workers need to do something totally mundane like drill a hole in a piece of metal, they need to bring the task to a union member to execute. Altogether, the union involvement made the summer quite different from my previous research summers in academic labs.

Polishing Problems

Polishing is a finishing operation, a treatment that affects just the outermost surface of the part you are working on. There are all sorts of different types of these finishes, and Zahner offers a wide variety to clients to use on their jobs. Ever since Anish Kapoor’s Cloud-gate, the mirror finish has become increasingly sought after. A mirror finish is achieved by polishing with increasingly fine abrasives until there are no remaining surface imperfections to disrupt the reflection of light. This can be done with sandpaper, with buffing wheels, or even with a rag and some rouge (polishing compound). The core issue with mirror polishing is that the process is extremely time-intensive. The work is only partly dependent on skill, so even expert metalworkers will still take a long time to convert a given surface area into a mirror finish. This time requirement leads to numerous other problems.

The typical tools used for polishing are pneumatically powered orbital sanders and rotary tools. The vibrations from these tools are not ordinarily a problem, but any worker involved in polishing work for a duration of time has increased risk of inducing nerve damage in their arms. One construction foreman I spoke to at Zahner mentioned that his arms had gone numb over time because of polishing. Then, there’s the issue that polishing removes metal particulates from the surface. Even with proper ventilation, metal dust is going to get everywhere. Proper PPE such as the appropriate respirator help to reduce the ingestion of metal dust, but metal dust is slow to leave the body. Regulations exist to limit the amount of dust ingested over time for metal workers. Past a certain point, workers have to stop polishing or they will accumulate toxic levels of the dust. Of course, all of these conditions are magnified if proper safety procedures aren’t followed. The final major issue with the polishing operation is that due to the major time cost, the work is notorious among fabricators. Companies have gone bankrupt due to inability to finish mirror polishing work. The work is difficult to bid for because it can take such a long time. Much of the time, the mirror polishing contracts will be given to the lowest bidders who do not factor long term worker health and safety into the bid. Mirror polishing is a common request for metal fabricators like Zahner, but increasingly they are failing to secure such time intensive contracts.

Overview of robotic polishing workspace

The R&D goal for the summer was to allow Zahner to bid competitively on these mirror finishing jobs without sacrificing workers. However, the union involvement immediately creates tension with this goal. The labor union certainly does not want their work to be handed over to the domain of robots. On the other hand, manual labor is failing to competitively execute the task, so the jobs are being lost anyways. Counter-intuitively, if robots could handle the brutal work of polishing, the union workers would secure more work on the assembly and production sides of those same jobs. Therefore, both those who demand and those who supply the labor are aligned in common interests. We should also address that the union consists of many members and they definitely were not all unified in wanting this work to be automated. The division among the union members was mainly between the younger members and the more experienced. The young metalworkers were typically not thrilled to see a robot being used to do their work. They saw this as the robot displacing them from work that would otherwise be theirs. However, anyone in the shop who had spent a considerable part of their careers polishing was excited to see the work being handled by the robot. These were the workers who had experienced the adverse effects of the work, and they did not want to assign their younger subordinates to the same tasks they had to go through.

Learning by Hand

To automate the polishing work, I had to learn from the union members themselves. I consulted with various workers who showed me their steps for achieving a mirror finish, the tools they preferred to use, and even the motions of their hands. Each had different preferences and methods that they had refined over their careers. Over the course of the summer I spent time scripting these motions and sequences until we were ready for the robot to try them out on some scrap metal pieces from the shop.

The major robotics obstacle we had to address during this summer project was the issue of how a robot arm actually interacts with a physical object. In most robotic fabrication projects, the digital environment and the physical environment will be calibrated to be roughly similar within a few millimetres, using careful measurements and registration markers to make sure everything is where it should be. However, mirror finishing requires much tighter tolerance than stacking bricks or folding sheet metal. To actually achieve a mirror finish, the surface has to be completely free from any scratches or bumps. When a worker polishes metal by hand, the combination of pressure from the hand and maneuverability afforded by the wrist let the worker adjust to keep the tool against the surface constantly. For humans, there is barely any mental calculation required to maintain this pressure, it comes very naturally to us. However, an industrial robot in its factory default configuration does not have any sensors, so it cannot make these minute adjustments on its own. We need to add sensors to incorporate force feedback to the process.

Active Control Flange (ACF) designed by FerRobotics.

Force Feedback

Fortunately for us, an Austrian company called FerRobotics offers a range of tools meant to interface with industrial robots to add the type of responsiveness we were looking for. The tool we got from them was an active control flange, essentially a very nicely manufactured linear axis. This linear axis would be attached to the end of the robot arm to give a seventh axis of motion. The tool could be programmed to maintain a constant surface pressure, and then within an inch or so of tolerance, the axis would constantly adjust to ensure that desired surface pressure. This extra inch of tolerance, though it seems quite small, makes a world of difference in terms of achieving a mirror finish.

Readying for Deployment

After we finished installing this additional axis and scripting the robot to re-perform the workers’ polishing motions, we did some tests to verify that the robot could reach the same level of mirror finish as a human. Part of the procedure that required lots of iteration to get right was the part where the robot needed to “refill” on polishing compound. Every few passes with the buffing wheel, the compound would run out, so we needed the robot to reach over and hold the wheel against the block of compound to reload. This end-effector assembly also required more wiring coordination than usual considering we had multiple air tubes and electrical cords to co-ordinate. The wires and tubes had to be fixed to the stationary points of the robot arm with enough slack between the connection points that the cords could never get twisted up or wrapped around the robot. After many iterations, we used a surface roughness meter to verify that we had achieved a true mirror finish. In the end the project was a success, and the scripts are currently in deployment in their first real project for Zahner where the robot is doing the polishing on a large sculpture. It’s interesting to note that this project had a remarkably quick turnaround from research to deployment. Mirror polishing contracts are quire lucrative, so the cost of a robotic polishing setup could quickly pay for itself with the contracts it would bring in.

Using a Surface Roughness Meter to check evenness of surface finish.

Patterning Mirrors

The more exciting part of automation work is when the automation enables new possibilities to emerge. At the end of the summer, we had some leftover time to do some more experimental work to see what the robot could do now that it was equipped with force feedback. The robot was suited to address many of the shortcomings of human labor in the typical mirror polishing application since it could work nonstop and without the same health consequences. Once we addressed the lack of force feedback, we unlocked the ability of the robot to perform highly precise motions with the tools. When making a mirror, theoretically it doesn’t matter what motions you perform. Criss-crotch hatching or zig-zagging are the fastest ways to contact the whole surface evenly, but really the same result could be achieved with any pattern of motions. However, since the robot can follow a precise path, we can intentionally mis-use the polishing operation to “draw” in the surface. Artists sometimes do this in metalworking, but at the scale of architectural metalworking, a human could never reliably pattern pieces at a large scale.

Now, it’s not that patterns in mirrors don’t exist already. They definitely do, but they are typically done with acid-etching of stencils on flat sheets. The constraint that the mirror finish has to be applied to the flat sheet limits the potential output. Something with double curvature like Cloud-gate would not be possible to do this way, since the sheets would have to be formed after being finished, which would likely ruin the finish. For the robot, the shape of the part or the complexity of the curvature is inconsequential. We did all of our tests on curved sheets as proof-of-concept, and had no issue imparting precise patterns on the curved forms. In this way, the addition of the robot not only relieves the metalworkers from one of their more time consuming and unhealthy tasks, but unlocks new sequences of assembly that expand the capabilities of metalworkers.

All Images property of Danny Griffin, taken while working under A. Zahner Company.