01

The task looked repetitive, but the exceptions mattered

The fictional assembly business Linden Devices wanted to assist operators with repetitive component positioning. Part presentation varied, the workstation space was limited, and experienced operators routinely corrected small alignment issues by feel. The team initially underestimated how much of the task depended on context.

We mapped the full sequence: presenting a part, checking its identity, positioning it, verifying alignment, and responding to exceptions. The pilot focused on the steps with repeatable inputs, while ambiguous or damaged components remained with a person.

02

A vision system with an explicit rejection path

The vision pipeline estimated part identity and pose, but only within the conditions covered by the evaluation set. Low confidence did not produce a best guess followed by motion. It produced a request for operator intervention.

The mechanical and safety design remained the responsibility of qualified specialists. Software evaluation was one input into the pilot review, not a substitute for risk assessment, protective measures, or validation of the workstation.

03

Measure the whole workstation

A robot’s motion time alone is an incomplete measure. We counted part presentation, operator corrections, recovery from a rejected observation, and the time required to switch component types. The assisted workflow had to be useful under ordinary variation, not only during a rehearsed demonstration.

Operators helped define the controls and the exception messages. A clear message such as “part identity not confirmed” was more useful than a model-confidence value without a suggested action.