Eight processes, one hand-built rig, and the robot time we did not buy

Testing candidate processes on a manual fixture instead of on production equipment.

Eight candidate joining processes were compared on a single hand-built manual rig instead of on production equipment because the question at that stage was which process worked well enough to deserve real investment. How fast or how repeatably a chosen process could run was a later question, and only that later question would have justified automated equipment. Booking robot time before the first question was answered would have meant spending real money to speed up a decision that had not yet been made.

A test kitchen before the batch line

A restaurant developing a new dish tries several different approaches on an ordinary stovetop before committing any of them to the large batch equipment the kitchen runs during service. The batch equipment is built to repeat one settled recipe reliably at scale, and tying it up with an afternoon of eight competing recipes would take it away from the work it was bought for. A kitchen that insisted on trying every candidate on the batch equipment would either slow the whole kitchen down or shrink its list of candidates to whatever the equipment had spare time for, and either way the final dish would be worse. A stovetop trial can also be abandoned halfway the moment a recipe is clearly failing, at almost no cost beyond the ingredients used.

The hand-built joining rig played the stovetop's role: cheap, flexible, and indifferent to production speed, leaving the equipment built for speed to carry on with its own job while the comparison ran elsewhere.

Holding the part still and letting the process vary

A welding robot is optimised for consistency and speed once a process has been chosen, repeating one proven motion thousands of times with a precision no hand can match. That optimisation is close to useless while finding out which process deserves commitment, because a cell programmed for one process cannot easily be reprogrammed eight ways in an afternoon. A hand-built rig holds the parts in a fixed, repeatable position and leaves the joining method free to change, so the process is the only thing varying from one test to the next. This is the ordinary design-of-experiments discipline of holding every variable steady except the one under test, and it applies to joining processes as directly as to anything else with several plausible options and no obvious front-runner.

The rig's crudeness helped. Anything more elaborate would have tempted someone to lock in a single process early to justify the effort already spent on the fixture, a bias a cheap, disposable rig is immune to.

Fair positioning is half of a fair comparison, and the other half is fair measurement. Every candidate was judged against the same set of measurements on the same kind of part, so that any difference in the results could be traced to the process itself. A trial in which one option is examined more carefully than the others quietly favours it before a single result is in, and the easiest option to examine carefully is usually the one somebody already prefers.

Why the ninth candidate was nearly free

A dedicated comparison rig also changes the economics of the evaluation. Once built, it costs nothing further and makes no demands on the production schedule, while every hour borrowed on a shared robot is an hour taken from the work it was booked for. It changes what gets tested as well. With the rig in place, adding a ninth process to the list cost only the parts and an afternoon, so more candidates received a fair look before a choice was made. An evaluation where each extra option must justify a large expense tends to narrow its shortlist early, and early narrowing is where better answers get missed.

The rig's job ends once a process is chosen. Manually fixtured trials rarely reproduce the positioning repeatability, cycle time and consistency of a robot, and a process that looked promising on the rig can behave differently on the equipment meant to build real parts at volume. The rig produces a well-informed shortlist, and confirming the winner on the production equipment is the discipline the next article in this set describes. Treating the rig's results as final is the easiest way to undo the value of the comparison.

My key error with this

When the consumable stopped arriving, I assumed the sensible move was to find the closest available substitute and confirm that it worked, which is a reasonable-sounding plan that quietly contains its own answer, because a test designed to confirm one option will almost always confirm it. I ran exactly that test, it passed, and I was ready to move on. The reason I did not is that somebody asked what the alternatives would have scored, and I had no answer, because I had never given them a chance to compete. Setting up the rig properly so that eight candidate methods ran against the same measurements took considerably longer than confirming my first choice had, and it produced a different winner. What replaced the belief is that a test which can only return the answer you already expect is not evidence, it is a formality, and the cost of finding that out on a production line is much higher than the cost of building the rig.