There is something to put away beneath a cart. The team can see it. It knows where it belongs. But it cannot move the cart.
There is a peculiar discomfort to this kind of standstill. The goal and procedure have supposedly been shared, yet the work does not advance. A little more intelligent thinking seems as though it ought to solve the problem. Perhaps what is needed, though, is not another thought but someone with a different capability.
A robotics study by Peiyan Li and colleagues brings that act of calling into collaboration itself. Who will work is not decided once at the beginning. The working team changes its own composition along the way. Intelligence finds a slightly different place to sit.
Changing the roster mid-task
The researchers used a pool of five robots with different capabilities and compared approaches to indoor tidying: keeping the original team, receiving help dispatched from outside, and having the team anticipate a shortfall and recruit a helper. There were 20 physical trials for each approach, 60 in all. Success rates were 45%, 55%, and 90%, respectively. [1]
This was not a contest in which a free-form team defeated an all-capable one. The starting configurations were designed to encounter capability gaps. “Why not provide every robot from the start?” is a reasonable response, but it leaves another question: how much space and how many resources may the team use?
The interesting part is that the story does not end when help arrives. In the proactive design, roles and the state of the work are coordinated before arrival. Capability reaching the workplace and capability beginning to work are not treated as the same event.
The reinforcements have arrived. Then more time passes while someone explains what they were needed for. It may be a familiar scene in a human workplace. Of course, this experiment did not measure human organizations. Even so, the boundary it illuminates feels close to home.
Nor does the implementation let robots imagine unlimited abilities. Each has a list of available skills, from which a language model selects the next action. The layer executing movement is distinct from the layer translating the work into language and coordinating with others. There is a boundary between free conversation and operations that can actually be performed.
This is what distinguishes it from simply letting robots chat. What is said connects to the next task, and the task’s result becomes material for further discussion. Fluency alone will not move the cart. A strong arm will not tidy the room either if it cannot reach the place that needs it.
The system was also evaluated in simulation across 10 layouts and 720 task scenarios. Yet perception and action there include idealizations. More rooms do not mean that every difficulty of reality has also increased. The large number should not hide the small number of physical trials.
A design that can say, “Something is missing”
The word autonomy tends to bring along an image of someone completing a task alone. Asking nobody. Borrowing no hands. Never stopping. The harder we pursue this picture of excellence, the more the moment of seeking help can look like a performance defect.
What follows is a reflection prompted by the study.
If autonomy means choosing the next action in relation to a goal, rather than finishing alone, calling someone else is no longer a failure of autonomy. It becomes an action that handles the boundary of one’s own capability. Being able to do everything and knowing what is missing are different abilities.
Applied as a question to human work, this might first change the meaning of escalation. Report upward when you get stuck. On its own, that arrangement leaves one person overwhelmed with requests and others waiting. Distinguishing what has stopped, which capability could move it forward, and what can be prepared before help arrives makes asking for help a job in its own right.
But simply increasing the freedom to ask is not enough. The person called has other work. The same people may be called repeatedly and carry a disproportionate load. The robot pool in this experiment has neither human fatigue nor calendars to negotiate. Those differences must not disappear when we bring the question into a company.
A handover needs an entrance
Joining midway needs a small entrance: the goal, the current position, what has been tried, and the decisions still open. Sharing everything is not necessarily considerate. A long record can turn a request for help into a reading-comprehension project.
Yet handing over too narrow a request, “just do this bit,” can also prevent other problems noticed by the helper from returning to the team. A handover resembles passing a parcel, except that the recipient may change the parcel’s shape. To leave room for that, we might distinguish task instructions from a shared understanding of what counts as a good outcome.
The robot experiment is not a finished answer. Physical validation was limited to a small indoor environment; unreliable communications and larger, more complex groups remain future work. Extracting the success rate and concluding that teams should become fluid would be too quick.
What remains is a question about where our design of collaboration begins and ends. We think about assigning work. We think about collecting results. In between, when capability runs short, who moves the boundary?
Reporting that something cannot be done does not finish the job. Could we treat the arrival of someone who can do it, and the conditions that let them begin, as part of the same piece of work?
A team stopped in front of a cart does not always need a larger brain. It may need a slightly wider entrance through which a colleague can join.