For a robot sharing a chore with you, moving objects to the right place is only part of being useful. The robot also has to notice which part of the chore the person has taken, what has changed since it planned, and when to wait or choose another useful step. Otherwise, two workers can spend more time doing the same job than one person would spend alone.
Imagine putting away groceries together. You start sorting food that needs refrigeration while the robot carries shelf-stable items to a cupboard. If it reaches for the bag you have already emptied, it has lost track of your progress. If it moves a package into a cupboard you are opening, it may get in your way. This is an illustrative scene, not a reported robot trial. It shows why a shared chore needs more than an object-recognition model and a list of destinations.
Sharing a chore means dividing a changing job
The first decision is who can and should do each step. Some tasks can be split freely; others depend on order or require a capability one partner lacks. In the PARTNR research benchmark, the robot can rearrange objects while the simulated human can also perform actions such as washing dishes or operating an appliance. A good plan must respect that difference. Assigning the robot a step outside its available actions is not collaboration, however neat the written plan looks.
Then the plan has to keep up with reality. Suppose the person has already cleared the plates from a table. A robot acting on its original instruction could return to collect them, search for plates that are gone, or undo a finished step. A useful partner checks the current state, recognizes that the plates are handled, and takes another available task. The same adjustment matters if the person changes the destination, pauses midway, or asks the robot to leave one item alone.
PARTNR was built to probe these problems. Its official project page describes 100,000 natural-language tasks in simulated homes, with spatial, temporal, and different-agent capability constraints. The size of the collection makes it useful for varied planning tests; it is not a count of household robots working in physical homes.
Finishing the task can still leave the person doing extra work
A final “done” label misses the quality of the partnership. Did the robot take a useful share, or did the person complete nearly everything while the robot wandered? Did both partners handle the same object? Did the robot wait for access to a cupboard, or force the person to stop and make room? Those questions matter even when the room eventually looks tidy.
The PARTNR paper reports that its tested language-model planners struggled with coordination, task tracking, and recovery from errors. In its human-in-the-loop evaluation, a real participant controlled a person avatar in the simulator while working with another person or an AI-controlled partner. The AI partnerships took more simulation steps than a person working alone in that study. This is evidence about the tested planners and virtual tasks, not a measured delay you should expect from a robot in your kitchen.
The practical lesson is to look at the person’s workload alongside task completion. If a robot takes twenty actions but the resident must continually redirect it, the resident has acquired a supervision job. A better account of a shared chore records who completed each part, where work was duplicated, what interruptions the robot caused, and whether it changed course after the person acted.
A partner needs to recover when the scene changes
Housework rarely follows a fixed script. A person picks up a cup the robot was about to collect; a drawer is closed; an object is moved to a different counter. The robot should recognize the new state before repeating an obsolete action. If a pickup fails, it should establish whether the object is still there and whether the person has already dealt with it before retrying.
That is a planning and observation problem as well as a manipulation problem. The official PARTNR repository includes planner baselines and simulated perception and action components in Habitat. The benchmark can expose failures in how a planner follows the other agent and the task state. It does not by itself establish that a physical robot can sense every household change, handle fragile objects, or move safely around a resident.
When judging a home-robot demonstration, watch what happens after the person changes course. Can the robot drop a now-redundant step? Can it leave a human-only action to the person and take another useful job? Does it recover from its own failed action without making the person repeat work? A smooth solo pick-and-place clip answers a different question. Our dishwashing guide explains the separate difficulty of handling the dishes themselves. A helpful partner must solve both the physical task and the shared-work problem.