The Cost of the Next Attempt

Imagine a robot inserting a connector. It approaches, makes a small adjustment, pushes the two halves together, and releases its grip. The camera gets a clear shot. The demonstration is finished.

The work around that motion is harder to see.

Someone has to establish whether the connection is secure. The completed assembly has to leave the fixture. A fresh pair has to arrive in a known condition. If the attempt failed, there may be a bent pin, a displaced cable, or a part stuck somewhere it should not be.

All of that has to happen before the robot gets another useful attempt.

For a system intended to learn through practice, the cost of this surrounding work matters enormously. A robot can spend seconds attempting a task and minutes waiting for a person to make the next attempt possible.

At Defex, we think this is an important part of making robot learning useful: engineering the conditions in which a machine can acquire experience repeatedly, judge it reliably, and improve from it.

From seeing a problem to doing the work

We started with visual inspection on factory lines. A camera could help identify a defect. The next question was how a robot could perform the physical work itself.

That adds a different kind of difficulty.

A connector can look aligned and still resist insertion. A cable can pull a part away from its expected position. Two pieces can appear joined while the latch has not fully engaged. The robot needs information about what happens as it acts, and a way to establish whether the finished result is acceptable.

Think about plugging in a cable somewhere you cannot see clearly. You use the shape and position of the plug to get close. Once it touches the socket, resistance tells you something new. You might back off, change the angle slightly, or stop because pushing harder feels wrong.

That interaction contains information a picture taken before the movement cannot provide.

For a robot, turning that information into a useful skill requires several things to work together: sensing, control, a meaningful outcome check, and a repeatable source of practice. Improving any one of them helps. The complete cycle determines whether the machine can keep learning without constant attention.

An attempt needs a trustworthy ending

A learning system needs to know what counts as success.

For assembly, “the robot reached its target position” is an incomplete answer. The tool could reach the expected position while the part slips in its grip. A force increase could indicate successful seating, an obstruction, or a misaligned component. A camera could show the expected outline while missing an internal problem.

Each measurement answers a particular question. None should silently stand in for every other question.

In our proposed connector task, electrical contact and mechanical engagement need separate consideration. Depending on the connector, electrical contact can occur before the assembly is securely latched. A continuity check alone could therefore reward an incomplete operation.

This matters because a learner improves against the feedback it receives. If the acceptance test rewards the wrong outcome, more practice can reinforce the wrong behavior.

Our approach is to define the physical acceptance procedure alongside the task. The robot’s observations help it act; a separate outcome check establishes whether the resulting assembly meets the required conditions. When the available evidence is incomplete or inconsistent, the system needs an uncertain result that can be investigated.

“It probably worked” is a poor foundation for both learning and manufacturing.

The same principle extends beyond connectors. Closing a lid requires more than reaching a final wrist angle. Placing an object requires knowing that it stayed where it belongs. Different jobs need different checks, but the learning problem always includes the question of how success becomes observable.

Designing that check is part of designing the robot’s education.

The world has to be ready again

Returning the arm to its starting pose does not reset a physical task.

The object has changed. Two parts may now be joined. A failed attempt may have moved a fixture or damaged a component. Consumables may be depleted. The next trial may begin under different conditions even when the robot’s joint positions are identical.

For a practice system, reset means returning the whole operation to a verified ready condition.

This creates a straightforward engineering constraint. For a bench that performs its steps sequentially, an approximate trial rate is:

The times are measured in seconds per trial. Recovery is the average additional time spent dealing with interruptions, spread across all trials. A cell that overlaps operations needs a more detailed scheduling model, but this simple accounting makes the issue visible.

Consider an illustrative bench with a 20-second action, a 10-second result check, and a 90-second reset. Assume there are no additional interruptions. It can complete 30 trials an hour.

Making the robot’s movement twice as fast raises that to about 33 trials an hour. Reducing reset from 90 seconds to 10 seconds raises it to 90 trials an hour.

These are hypothetical numbers, not Defex measurements. They illustrate why the next attempt deserves its own engineering effort. Faster motion has limited value when most of the time is spent restoring the task.

There is also a labor constraint. If every reset requires a person, collecting more experience requires more of that person’s time. An unattended interval is useful only when preparation, replenishment, and later recovery are included in the accounting.

For our first workcell, we are planning around prepared carriers that hold the mating parts and restrain the cables. After an attempt and its checks, the carrier would move to an accepted or quarantine destination, and a fresh pair would advance.

That is a finite batch of prepared work. Loading it still takes effort, and eventually it needs replenishing. We want to measure those costs explicitly.

This approach also avoids making automatic separation and endless reuse of the same connector pair prerequisites for the first experiment. Repeated use can change a component’s condition. The system needs to know what material it is practicing on and when that material should leave the experiment.

Repetition has to produce evidence

A high trial count is useful only if we can understand what those trials mean.

An unsuccessful insertion can reveal a lot. The robot may have approached from the wrong angle, contacted the edge of the housing, or encountered an unexpected cable load. But the final label “failed” does not identify which of those happened.

We want each attempt to connect the starting conditions, commanded motion, contact measurements, physical result, and any human intervention. That makes it possible to distinguish a behavior that needs improvement from a displaced fixture, an unsuitable specimen, or a faulty check.

Failed attempts belong in this record. So do occasions when a person intervenes. Removing those events would make the system look more autonomous while obscuring where it actually needs work.

The practice setup also needs deliberate variation. Repeating an identical motion with an identical part can test repeatability. To investigate adaptation, we need defined changes and evaluation conditions that reveal whether a skill handles them.

For a first connector task, this means establishing an operating range and recording the variation within it. It also means reserving evaluation trials so that a candidate behavior is judged on more than the attempts used to develop it.

The objective is experience that supports a decision: what should change, whether the change helped, and where the resulting skill can be trusted.

What we are building first

Our first assembly workcell, A1, is intended to make this cycle concrete.

We have developed an inspection prototype. For A1, the assembly architecture and learning plan are specified; the next milestone is an instrumented connector bench. The practice and adaptation capabilities described here are what that program is intended to test.

We are starting with one fixtured connector family, then a related variant. It gives us a bounded physical operation, a result we can test, and a meaningful next task on which to measure reuse.

People still have essential work to do. They define the operation, select the tooling, establish limits, prepare the material, and agree on the acceptance procedure. Initial teaching can include demonstrations and corrections using the robot itself.

When we describe the direction as self-teaching, we mean that recorded attempts and feedback should help the system learn useful adjustments within that engineered setup. The amount of human work required is something we intend to measure.

The first comparison should be against a competent conventional solution. If a programmed sequence and suitable contact control perform the task reliably, that establishes a baseline. A learned behavior has to earn its place through better results or less adaptation effort.

This keeps the choice of learning method tied to a concrete problem. Demonstrations might help establish an initial skill. Corrections might address recurring failures. Further training might improve performance over the defined range of conditions. Each step needs evidence that it contributes something useful.

We also separate practice from production. During practice, candidate behaviors can be developed and evaluated. Production should run a specific qualified version with known operating limits and a fixed acceptance procedure. A new candidate earns deployment through evaluation.

That separation lets us investigate improvement while keeping the meaning of an accepted part consistent.

The second task is the important test

Getting the first task to work is necessary. It does not, by itself, establish that experience makes the next task easier.

A team can invest substantial effort in fixtures, programming, demonstrations, debugging, and testing to produce one successful installation. The next installation may require much of that work again.

Our larger ambition depends on changing that relationship.

Suppose the workcell qualifies connector A. We then introduce a related connector B with a different geometry. The question is how much of the first task’s knowledge and engineering can be reused, and how much new effort is needed to reach comparable quality and useful throughput.

The comparison needs to count the complete adaptation process:

• Tooling, fixture changes, and calibration.
• Programming, demonstrations, corrections, and training.
• Debugging, failed attempts, and damaged material.
• Validation and human intervention during operation.

Training time alone cannot answer the question. A behavior that trains quickly but requires days of fixture adjustment may offer little overall benefit. A high success rate can also conceal a slow cycle or frequent operator assistance.

For the first experiments, these are the measurements we want to make visible:

There is a useful consequence of measuring this way: reuse can come from several places. A motion skill may transfer. So may a fixture concept, a test method, or an understanding of a recurring failure. We should identify those contributions honestly rather than attributing every improvement to the learning algorithm.

The customer experiences the combined result: how much work it takes to get the next operation running reliably.

Making experience accumulate

Connector assembly gives us a place to start. The wider goal is robots whose useful capabilities become easier to extend as they gain experience.

That requires an entire process around the motion: material ready to work on, feedback during contact, a trustworthy result, another attempt, and an evaluation that shows whether anything improved.

There are difficult questions ahead. How much variation can a learned skill handle? How often will recovery still require a person? Which parts of the setup will transfer to the next job? The first workcell needs to produce evidence that answers them.

Our bet is that making practice measurable and repeatable will help turn robot learning into something a factory can use. Every qualified task should leave behind knowledge that reduces the effort of the next one.

That is the direction we are building toward at Defex: robots that can learn useful physical work, with experience that continues to matter after the first demonstration.