By Rafael Patron, president of the Scientific-Technical Committee, AIPIA
Supervised autonomy in robotics is usually described from the robot’s side: the machine does the repetitive work, and a person steps in when something goes wrong.
The harder design question sits on the human side, and many projects leave it until commissioning, when nobody has yet agreed who answers the robot when it gets stuck.
Picture a depalletising cell on a night shift. A pallet arrives wrapped in a reflective film the vision system has never seen, the gripper fails twice, and the line behind it begins to back up.
Somewhere a supervisor gets an alert. What happens next depends less on the AI model than on decisions taken months earlier.
Why the robot’s worst day is a design requirement
In a recent interview with Sam Francis for Robotics & Automation News, Plus One Robotics chief executive Erik Nieves argued that a robot should be judged on its worst day: how it behaves in front of something it has never encountered, rather than on a smooth demonstration.
He compared the role of human supervision to the pilots who stay in the cockpit while the autopilot flies, rarely needed but always available.
The comparison is useful because aviation did not leave the cockpit to chance. Who takes control, when, with which information and after what training is written down, rehearsed and audited.
If today’s AI still struggles to recognise the limits of its own competence, as Nieves put it, then the human backstop has to be engineered with the same care as the gripper and the conveyor. The cost of handling exceptions belongs in the business case alongside cycle time.
Mapping exceptions before the robot meets them
Start with an inventory. Any operations team can list the situations that disrupt a manual process today: damaged cartons, mislabelled parcels, items stuck together, a pallet pattern that changes without notice, a person walking into the wrong zone. Each one deserves a decision before go-live.
Some exceptions the robot may resolve alone, within limits: retry a pick from another angle, set an item aside in a reject lane, slow down.
Others must always reach a person, either because the consequence is costly or because the robot’s confidence cannot be trusted in that context.
Anything involving safety, product damage above a threshold, or a customer-specific rule usually belongs in the second group.
The boundary will move over time. With more data, robots ask for help less often, a point Nieves made in the same interview.
A well-run site reviews the list every few months and moves some exceptions from “always escalate” to “resolve and log”, which only works if every exception is recorded, including those the robot solved alone.
What a good call for help contains
A robot that asks for help with a bare error code turns the supervisor into a detective. A good request carries context: a camera image or short clip of the scene, what the robot tried and how many times, its own confidence estimate if the model provides one, and what will happen to the line if nobody answers.
It should also say how long the supervisor has before the cell stops or diverts the item.
That last element changes behaviour. A supervisor who knows that an exception can wait two minutes handles it very differently from one who sees every alert flashing with the same urgency.
The interface matters as much as the content. Supervisors who watch several cells, sometimes remotely, need requests ranked by urgency and grouped by type, so that a recurring problem shows up as a pattern rather than twenty separate interruptions.
Every decision the person takes should feed the exception log, the raw material for the next review.
Supervised autonomy needs a real supervisor role
Nieves expects supervised autonomy to be the main shift of the next five years, with each supervisor overseeing more robots as the machines resolve more exceptions themselves.
That makes the robot-to-supervisor ratio a planning variable, and it is easy to get wrong in both directions.
Too few supervisors and the line waits; too many and the role becomes idle monitoring, which erodes attention when the real exception arrives.
There is no reliable industry benchmark for that ratio, and any figure taken from another site will mislead.
A practical way to set it is to measure, during the pilot, how many requests each cell generates per hour, how long a person needs to resolve each type, and how the numbers change across shifts.
The ratio then follows from the site’s own data, and it can be revised as the exception list shrinks.
The skills are new as well. A good supervisor needs to read the robot’s signals, understand roughly why a vision model fails on reflective or deformed items, recognise when a recurring exception points to an upstream process problem, and know when to stop a cell.
This is closer to process engineering than to machine minding, and it deserves a defined role, training and a career path.
For organisations deploying AI systems in the European Union, Article 4 of the AI Act, Regulation (EU) 2024/1689 as amended by Regulation (EU) 2026/1744, also asks providers and deployers to take measures to develop the AI literacy of their staff, taking into account their knowledge, experience and the context of use.
Where automation engineers and AI professionals meet
Robotics & Automation News follows the question of trust in AI-driven robots largely through conversations with the people who build and deploy them, from warehouse picking to humanoids.
Its readers are the engineers and operations managers who will have to turn that trust into shift patterns, escalation rules and job descriptions.
Part of that work falls to a different profession. The people who tune confidence thresholds, design the logging behind each request for help and decide what the model may resolve alone are increasingly AI specialists working alongside automation engineers.
The Italian Association of Artificial Intelligence Professionals (AIPIA) brings together developers, data scientists, researchers and consultants who do this kind of work in organisations.
A useful reference here is the Italian standard UNI 11621-8:2026, published in April 2026, which describes twelve professional role profiles for artificial intelligence, from the Chief AI Officer to the AI Consultant and the Data Scientist.
Certification against the UNI 11621-8 standard is issued by an accredited third-party body under ISO/IEC 17024, while AIPIA offers training paths aligned with the profiles.
For a plant manager, the practical value lies in the vocabulary. When the AI side of a robotics project is described in terms of defined roles and responsibilities, it becomes easier to decide who owns the model’s behaviour, who owns the floor, and who signs off a change to the exception list.
Being honest about the limits
The survey Plus One Robotics commissioned, quoted in the same interview, shows why this matters beyond the factory gate.
Among Americans asked about humanoid robots, 26 percent named the lack of human judgement as their main concern, and roughly 58 percent of those uncomfortable with certain robot tasks said they would feel better if a trained person could monitor and intervene.
Discomfort ran highest for robots watching children, at 52 percent, followed by surgery at 46 percent and caring for elderly people at 39 percent. It is a company-commissioned survey about humanoids, but the message travels to warehouses.
Writing the limits down is the most direct answer to those concerns. A one-page description for each cell, stating what the robot handles alone, what always goes to a person and how quickly that person responds, does more than any claim of autonomy.
It also protects the operator commercially: a customer who knows that certain items are diverted for human handling will not read every diversion as a failure.
Nieves warned against selling home and care robots as autonomous before they can cope with unpredictable human settings. Warehouses have an advantage: they can be open about supervision without losing face.
A simple test for any site planning an AI-driven cell is to ask four people, before the contract is signed, to describe the robot’s worst day: the operations manager, the automation engineer, the AI specialist and the supervisor who will be on shift.
If their answers differ on who acts, with what information and within how long, the design work is not finished.
About the author: Rafael Patron is president of the Scientific-Technical Committee of AIPIA, the Italian Association of Artificial Intelligence Professionals, and chief AI officer at Intarget DMCC. He has more than 20 years of experience in technology, digital transformation and AI strategy. His work focuses on AI adoption, governance, professional standards and education. He is also an author and lecturer on artificial intelligence and digital technologies.
