200 Hours, Then One: the Gearbox Cell That Writes Its Own Insertion

The photograph is the entire argument in a single frame. A blue-jointed UR cobot is lowering a sun gear and carrier into the ring gear of an e-axle planetary set, seated in an aluminium housing, sitting inside a moulded foam tray. Nothing about the motion looks remarkable — and that is precisely the point. Producing that motion used to cost 200 hours of engineering.

The problem was never a trajectory

Take the geometry seriously and the difficulty becomes obvious.

An e-axle planetary gearbox has a ring gear fixed in the housing with internal helical teeth, three planet gears that rotate independently on their carrier pins, and a sun gear that drives them. Before assembly, each planet sits at an arbitrary angular position, and their positions relative to one another are also arbitrary — the parts shift freely in the tray, and their weight, size and shape vary within tolerance.

Because the teeth are helical, insertion is not a straight push. The helix acts like a screw thread: each planet has to rotate while it travels axially, at a rate set by the instantaneous tooth contact. Three planets must engage the ring gear simultaneously, and one jam tilts the carrier and stops the other two. All of it happens inside the housing, out of sight.

This is why conventional robot programming fails here. A taught path is a fixed sequence of joint positions; the task is a phase search under contact with three coupled unknowns that reshuffle on every part. You cannot enumerate the cases, so you cannot pre-program them. Engineers responded the only way available: by hand-tuning a cell until it worked for the parts in front of them, then rebuilding it when the batch changed.

The number that matters, and the number that misleads

The headline figure comes from AICA’s own project record with Gebert Rüf Stiftung, the Swiss foundation that funded the company’s early work: Schaeffler “used to spend 200 hours programming a robot for gearbox insertion. Instead, with our solution, they are training a robot in one hour.”

That is a 200-fold reduction, and it is the number that travels. It is also the number most likely to be misused.

The International Federation of Robotics published a case study on the same relationship in August 2025, after the technology moved into series production. Its figure for the industrialised project is more than 30 percent reduction of development time — along with no need for dedicated robot-programming expertise, and a hardware-agnostic implementation.

Both numbers are official. Both are true. They measure different things:

• 200 hours to 1 hour is the per-batch re-teaching loop — the part of the work where an engineer stands at the cell with a teach pendant, nudging approach vectors until the teeth stop jamming.
• 30 percent is the whole engineering package: fixture design, safety, cycle-time tuning, integration, validation, documentation.

If you build a business case on 200×, you will be wrong by an order of magnitude. Roughly a fifth of the engineering effort disappears; it happens to be the fifth that everyone hated.

What “about 100 trials” actually demands

The detail that should shape a project plan is buried in the same record: the learning approach needs “about 100 assembly and disassembly steps in the gearbox insertion use case.”

One hundred. Not a hundred thousand. Which tells you something structural about what is being learned: the policy is not discovering gearbox assembly from scratch. It is discovering the residual — the specific clearance, compliance and friction of this cell, with these parts, on this fixture. The prior comes from engineered force control; the learning is a thin adaptive layer on top.

Two consequences follow.

First, a buyer who expects reinforcement learning to compensate for a poor force loop will be disappointed. The impedance behaviour is designed, not learned. If the controller cannot hold a stable compliant contact, no amount of policy search rescues it.

Second, 100 trials is an operational requirement before it is a software one. The cell must run roughly 100 unsupervised insert-and-extract cycles at commissioning — overnight, unattended, without destroying parts. That means:

• 100 good housings and 100 gear sets available as learning stock, budgeted as consumables;
• a defined disposition for those 100 units — scrap, rework, or re-test;
• a per-trial record showing what the policy converged to, for the PPAP file;
• and a cell that can be left running with nobody watching it.

None of that is in the software licence. All of it lands on the integrator and the plant.

The success criterion is the quality system

The mechanism is described precisely in the IFR material: the robot “independently and autonomously trained the optimum assembly parameters based on previously defined framework conditions and success criteria.”

Read that last clause again. An optimiser maximises whatever you declare. Declare “fully seated” and you will get gears forced into place, some of them damaged, at a cycle time nobody planned. Declare nothing and you have automated the production of scrap with excellent repeatability.

So the specification becomes the product. In practice the success criteria have to include, at minimum:

• a peak axial force envelope, high and low;
• a torque-versus-rotation signature during engagement, not just a final torque;
• insertion depth with a tolerance band;
• a radial force limit, which is the signature of a tilted carrier;
• a cycle-time ceiling, or the policy will find an elegant solution that takes four minutes.

Whoever writes that list owns quality on that cell. It is a process-engineering deliverable, and it is the part of the project that cannot be bought.

Hardware-agnostic changes where you spend the money

Schaeffler’s stated requirement was explicit: the ability to “switch between different robot brands without a lot of hassle.” AICA’s architecture delivers this because the policy is expressed in task and force space rather than joint space — the same reason it transfers across a cobot and an industrial robot.

That has a purchasing consequence that is easy to miss. If insertion quality is decided by a force policy rather than by taught joint positions, then the arm’s repeatability specification matters less, and three other things matter more:

• the wrist force-torque sensor — resolution, bandwidth and drift;
• the gripper — its compliance and, critically, its centring repeatability;
• the fixture — stiffness and kinematic constraint, because the policy will exploit any slack you leave it.

Value migrates from the six-axis arm to the wrist. For component suppliers, that is the line of the BOM to be selling into.

Why the first series line was in Greater China

The series-production use case identified during industrialisation was a helical planetary gearbox for an e-axle drive, in the e-mobility field, in Greater China. The IFR frames the driver as a shrinking skilled workforce, and that is genuinely part of it.

But the arithmetic of the savings points somewhere more specific. A 200-hour saving is a one-off; a 1-hour re-teach is recurring. The benefit scales with the number of batches and design changes per year, not with the wage rate. An e-axle line running frequent variant changes — new ratios, new housings, new customers, short model cycles — compounds the recurring saving every few months. That, more than labour cost, is a defensible reason for the technology to reach series production in the market with the fastest variant cadence.

The uncomfortable read for component makers

If assembly know-how becomes a purchasable software module, precision assembly stops being an internal moat and becomes an input.

For a reducer or gearbox manufacturer, that cuts both ways. The barrier to automating a difficult insertion falls — for you, and for the six competitors behind you. What does not fall is everything the policy has to cope with: part-to-part consistency, fixture and metrology design, and the ability to write a success criterion tight enough to protect your customer.

There is a second-order effect worth noting. A policy that converges in 100 trials on good parts will take longer, and fail more often, on inconsistent ones. The technology quietly raises the value of tolerance control — which is good news for anyone actually selling precision, and bad news for anyone selling the appearance of it.

What to ask before signing

Five questions that separate a reference story from a project plan:

1. Where do the learning parts come from? Who pays for the ~100 units consumed at commissioning, and what is their disposition?
2. What is the declared success criterion, in force, torque and depth numbers — and who signs it? If the answer is a verbal description, the quality system is a verbal description.
3. What is the force loop bandwidth and sensor resolution? Then: what happens to the learned policy if I change the gripper?
4. Show me the per-batch re-teach time on my part family, not on the reference gearbox.
5. What does the cell do when the policy fails at hour six of a night shift? Recovery behaviour, not insertion behaviour, is what determines whether the line runs.

Related reading

• Harmonic drive torque sensing: where the accuracy actually lives — the wrist-level sensing this kind of policy depends on.
• Robot cell three sums: what to check before you buy — the cost arithmetic that decides whether a 200-hour saving is worth a project.
• Hollow-cup motor winding: the wire is the part — the other assembly problem that resists pre-programming.
• Tesla’s Optimus line: 1M a year, 20k a week — what takt time does to assembly economics.
• Robots building robots: the takt-time argument — why assembly labour, not assembly code, sets the ceiling.

Sources: the Schaeffler use case, the 200-hour and 1-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, the hardware-agnostic requirement, the long-term qualification test and the Greater China e-axle use case are from the IFR case study published 6 August 2025; the Automate 2026 UR7e demonstration and the company background (founded 2019 in Lausanne as an EPFL LASA spin-off) are from Gear Technology, August 2026. The reading of the two time figures, the success-criterion argument, the wrist-value-shift reading and the purchasing checklist are Latest Robot News editorial. Photo: AICA.