A robot’s capacity includes the people who support it

A robot works independently for an hour. Eventually it gets stuck on packaging positioned differently from before. Someone helps briefly, rearranges the material and restarts the process. How much work did the robot take over during that hour? The answer also depends on what the person providing help could do in the meantime and how quickly they had to be available.

This question becomes relevant to hospitals, medical technology companies and other organisations as physical AI moves towards a place in their operations. A demonstration shows a capability. Planning also requires an understanding of how that capability behaves across a working day with changing conditions and limited human support. The short interruptions deserve particular attention: they can determine how the entire process needs to be organised.

Chelsea Finn's Y Combinator talk, published on 12 August 2026, prompted this reflection. Her account of learning from experience leads to an operational question that I develop here: how should a robot's capacity be described together with the people who support it? The following proposals offer a way to think about the organisation of physical work. [1]

Support serves different purposes

Physical Intelligence’s November 2025 π*0.6 preprint studies laundry, espresso and box assembly. Throughput on difficult laundry and espresso tasks more than doubled from the preceding training stage. Human feedback, corrections and resetting the experimental environment remained necessary. Clinical uses were not studied. [2]

For operational assessment, I would divide human support by purpose. A person may demonstrate a new process, assess a situation, correct an unsuccessful movement or prepare the environment for another attempt. These activities require different skills. They also occur at different times. An initial teaching task and immediate assistance required during every shift have different implications for staffing.

My inference is that a single statement about “autonomy” will rarely provide enough information for a purchasing decision. It should be clear which support activities are included in the reported period and which lie outside it. This can be established through straightforward questions. Who prepared the materials before the demonstration began? What happens between two runs? Who decides what to do in an unclear situation? How long does work take to resume afterwards?

Measure the interruption from the person's perspective

Consider a fictional example from a hospital group's administration. A robot sorts sealed office-supply packages into transport containers. The task involves neither patients nor medical products. An employee is also supervising another process. Occasionally, she has to walk over to the robot, align a container and then return to her previous task.

A machine-only account might record just the time when the gripper stops moving. From the employee's perspective, walking over, understanding the situation and resuming the other task also take time. For useful planning, I would consider these durations together. Measurement should remain simple enough to change the observed process as little as possible. A small, time-limited investigation can already reveal which elements have been overlooked.

The distribution of interruptions also matters. Ten evenly spaced requests for help can have different consequences from ten simultaneous interruptions across several devices. Even with the same total duration, the second situation may require more people who are immediately available. A monthly average reveals little about whether the required support can be reached at a particular moment. This belongs in the design of collaboration between people and machines.

The environment belongs in the performance description

The 2024 DROID conference paper describes demonstrations from 564 scenes using a shared robot platform. It examines environmental diversity in learning manipulation tasks. This provides no assurance that a particular system can handle every new workplace without further assessment. [3]

I would derive a simple rule for organisations: every performance statement should include a description of its environment. That means the objects, their possible arrangement, the spatial conditions and the changes allowed while work proceeds. The description should also identify what was deliberately held constant during the assessment. Another team can then judge how closely its own operating environment resembles the one tested.

In the fictional example, a supplier might change its packaging. Its surface could be different even when the contents and item number remain the same. This is initially a change in operating conditions. A responsible person should be able to determine whether it remains within the tested description and which reassessment is needed. An expanding record of experience helps, provided it identifies such changes clearly.

Plan learning and ongoing operations separately

When support helps improve a robot, it may create value beyond the current run. I would initially treat that value as an expectation. The team should later be able to check whether the particular difficulty occurs less often or becomes easier to resolve. Otherwise, continuing supervision can always be described as an investment in future autonomy, without the expectation ever being tested.

That requires a traceable connection between changes and observations. Was a new model version introduced? Were the materials arranged differently? Did a more experienced person take over? When several things change at once, the cause of an improvement remains open. Preserving that uncertainty helps subsequent planning. A better process can be valuable even while its technical and organisational contributions remain difficult to separate.

The parties should also agree when learning during operations is allowed and which changes require assessment beforehand. The concrete decision depends on the use, its possible consequences and the responsible specialists. A demonstration of household tasks provides no general permission for deployment elsewhere. In medical environments especially, the requirements of the actual intended use need to be established separately.

Make a capacity statement that can support planning

For the next discussion with a supplier, I would request a description that brings successful work, human time and operating conditions together. How much was completed, over which period? What support was required? Which interruptions were observed? What responsibilities remain with the operator? These questions clarify which resources must actually be available for the proposed deployment.

From my perspective as a founder, this also creates a fair basis for collaboration. A supplier can demonstrate progress more precisely. An organisation can decide whether the remaining workload fits its operations. The people who sustain those operations are involved early. Through that work, an impressive movement becomes a capability that others can responsibly include in their plans.