The Invisible Hours: Where a Robot Gets Its Capability

A new industrial robot, uncrated, is a set of metal arms. It does not grasp, it does not transfer, it does not perform a process. What gives it capability is a laptop at a commissioning station — and a person who spends the next several weeks sitting in front of it.

That observation comes from a short post by a Chinese commissioning engineer, published under the title 给机器人注入灵魂 — “giving the robot a soul”. It is not a technical document. It circulated among automation people because it names something the industry prices badly: the work between delivery and production, which no quotation ever itemises and every schedule compresses first.

What the work actually consists of

The post lists it in the order it happens, and the list is worth taking literally, because each item is a place where a project loses weeks:

• Writing and editing the program. On an ABB cell that means RAPID — motion instructions, logic, interrupts, and the module structure that decides whether the next engineer can read it.
• Calibrating coordinate frames. Tool centre point, workpiece frame, base frame. Get the TCP wrong by two millimetres and every taught point inherits the error, silently.
• Verifying IO. Every handshake with the gripper, the safety plc, the conveyor, the vision trigger. A single inverted signal can present as an intermittent fault that takes days to find.
• Teaching points, then re-teaching them after the fixture moves, the part changes, or the approach angle turns out to collide.
• Avoiding singularities. Wrist and shoulder singularities are not errors the controller reports in advance; they are regions where the arm’s behaviour degrades, and the fix is usually to redesign the path or move the fixture.
• Tuning cycle time to takt. A program that works is not a program that makes rate. Getting from “it runs” to “it runs at the number in the business case” is its own project.
• Low-speed dry runs, then alarm handling, then the fault that only appears on the third shift.

None of this appears in a robot’s datasheet. All of it determines whether the cell pays back.

Why it stays invisible

Three structural reasons, and they compound.

It is not in the robot’s price. It sits in the integrator’s hours, spread across a project budget as “engineering” or “integration”, rarely as a number anyone can audit. AICA, the Swiss force-control software company, puts the extreme version of this in its own marketing: that programming can approach twice the price of the hardware alone during integration. That is a vendor’s framing and should be treated as one, but the direction is familiar to anyone who has signed a cell acceptance.

Demos show the result, not the iteration. A trade-show video is the tenth version of a motion. The nine versions that collided, alarmed, or ran at half the required rate are the actual project.

It does not transfer. Every cell re-learns its own frames, its own tolerances, its own quirks. Two identical cells in two plants still require two commissioning efforts. This is precisely the property that software vendors are now attacking — and the reason the numbers coming out of that work are so large.

The one number that exists

There is very little public data on commissioning hours, which is itself telling. One figure does exist, because a funder required it: in AICA’s project record with the Swiss foundation Gebert Rüf Stiftung, Schaeffler reported that it “used to spend 200 hours programming a robot for gearbox insertion” and that with the new approach it trains a robot in one hour.

Two hundred hours, for one task, on one cell. Not a whole line — one insertion. That is the shape of the invisible cost, and it explains why the industrial response to it has been to try to automate the search rather than the motion.

The International Federation of Robotics, reporting on the same relationship once it reached series production, gives the more conservative project-level number: more than 30 percent reduction in development time, plus no need for dedicated robot-programming specialists. Both figures are official. They measure different scopes, and a buyer should keep them apart: the 200 hours is the re-teaching loop, the 30 percent is the engineering package around it.

What is being automated, and what is not

Force-controlled assembly with learned parameters attacks exactly the part of commissioning that is pure trial and error: the approach vector, the contact search, the insertion that has to be discovered rather than calculated. Where that works, the teach pendant loop collapses.

What it does not touch is judgement. Three things stay human, and they are the ones that decide quality:

1. The success criterion. A learned policy optimises whatever you declare. Declare “inserted” and you get insertion at any cost, including damaged parts. The force envelope, the torque signature, the depth band — those are written by an engineer, and they become the quality system.
2. The fixture and the part. A policy that converges in about a hundred trials on good parts will take longer and fail more often on inconsistent ones. Commissioning time is a function of part quality, and no software changes that.
3. Diagnosis. When a cell stops at 3 a.m., the question is which of forty subsystems changed. That is not a tuning problem.

So the honest reading of the 200-to-1 number is not that commissioning disappears. It is that commissioning shifts from finding the motion to specifying and proving it — a smaller, more senior, more document-heavy job.

What to put in the integration contract

If commissioning is the unbudgeted tail of automation projects, the fix is contractual, not technical. Six clauses that change the outcome:

• Cycle time measured at takt, with the acceptance run defined in advance. How many hours, on whose parts, with whose operator, and what counts as a stop. “Meets the required cycle time” without this is an argument, not a spec.
• A reach and singularity study signed off before mechanical freeze. Singularity problems found on site are layout problems, and layout problems are change orders.
• The IO list and naming convention delivered as-built. Otherwise every future change routes back through the original integrator, at their rate, forever.
• Source code and project archive ownership, with a licence that survives the relationship. A cell you cannot modify is a cell you cannot improve.
• Commissioning hours as a priced line item with a rate and a cap. Absorbed into “integration”, they are the first thing squeezed and the last thing measured.
• A defined acceptance for fault recovery, plus who is on call for the first 90 days. Recovery behaviour, not nominal behaviour, is what decides whether the line runs.

The part worth keeping

The post ends on a line that translates cleanly: the machine is hardware; human thinking is the actual content of automation.

That is not sentiment. It is a cost statement. The metal is commoditising — the Big Four now hold about 38 percent of the global market — while the hours that turn metal into output remain stubbornly local, stubbornly manual, and almost entirely unmeasured. The companies that win the next decade of automation will be the ones that can see those hours, price them, and then remove them deliberately rather than discovering them at the end of a project.

Related reading

• 200 hours to 1 hour: the gearbox cell that writes its own insertion — the case study behind the number, and what 100 trials actually cost.
• Robot cell three sums: what to check before you buy — the procurement arithmetic that should include commissioning.
• The Big Four at 38%: a repricing — why the hardware is commoditising while the engineering is not.
• Hollow-cup motor winding: the wire is the part — another process where the difficulty lives in the setup, not the machine.
• Tesla’s Optimus line: takt time as the real constraint — what happens when a line is designed around rate from the start.

Sources: the commissioning task list and the closing observation are adapted from the Chinese post 给机器人注入灵魂 by 机器人调试手记; the 200-hour and one-hour figures and the ~100 trial count are from AICA’s project record with Gebert Rüf Stiftung (GRS-016/21); the 30 percent development-time reduction and the removal of dedicated programming specialists are from the IFR case study published 6 August 2025; the “up to twice the hardware price” framing is AICA’s own published claim. The contract clauses, the reading of the two time figures, and the shift-from-search-to-specification argument are Latest Robot News editorial.