Who is actually controlling your robot?

Start with a task you can observe

A robot crosses a room, picks up a shirt and places it in a basket. The movement is visible; the decision process is much harder to see. Someone might be steering continuously, choosing a destination, approving a difficult step or doing nothing after the initial instruction. A useful autonomy claim tells you which of those situations occurred and what the robot was expected to accomplish.

Begin with a concrete question: can this model carry my empty cup from this table to that counter, under these conditions, without another instruction? Keep the object, environment and completion rule specific. A machine can handle one repeatable task well while needing help with another. Describing the whole product as autonomous hides that variation and makes comparisons between very different robots surprisingly unhelpful.

Separate the kinds of human involvement

Remote operation can coexist with sophisticated software. A person may choose where a walking robot goes while its onboard controller manages balance. In that situation, the robot is solving a difficult control problem, but the person still decides its route. Microduck's public software repository provides a concrete example: gamepad input and neural movement policies are documented together. A controller in the operator's hands does not make the robot's balance imaginary.

For a product comparison, record the control arrangement in ordinary language. These descriptions are more informative than a single score, and a product may use several of them during the same session.

• Direct operation: a person continuously commands movement.
• Task selection: a person chooses a goal and the robot executes a defined behavior.
• Supervised operation: a person monitors progress or approves decisions.
• Exception assistance: a person intervenes after the robot gets stuck.

Read assistance beside the sales promise

On the NEO order page checked on September 6, 2026, 1X describes basic autonomy for early owners and scheduled remote expert supervision for unfamiliar complex tasks. That is a material part of the offer. It gives a prospective owner a much better question than whether the robot uses AI: which chores can it finish today, and when would an expert be involved? These are the manufacturer's descriptions, not results from a Robots at Home home trial.

Before buying any assisted system, establish how assistance begins, whether you approve it, what the operator can access and what happens when help is unavailable. Ask for the expected intervention frequency for your intended task. An occasional recovery and a person guiding every minute imply very different ownership experiences, even if the edited footage looks similar.

Map the computer as well as the operator

A robot can make decisions automatically while sending information to a remote computer. It can also run code locally while a human controls it. Local processing, cloud processing and human involvement therefore deserve separate fields. A Wi-Fi connection proves none of them by itself: the connection might carry updates, video, commands or the entire perception service.

Reachy Mini's documentation distinguishes a wireless model with onboard computing from a Lite model connected to a computer. Its app documentation also describes configurable software integrations. That means you need to examine the selected model and application, not infer every dependency from the robot's name. List the camera, microphone, controller and any external service, then ask what stops working if each connection disappears.

Look for the parts a short clip cannot show

A useful demonstration states the starting conditions, displays the complete task and explains interventions. You also want to know how many attempts were made, what counted as success and whether the scene changed between runs. Moving one familiar object from a fixed position is a different challenge from finding unfamiliar objects in a rearranged room. Both can be useful, provided the description matches the experiment.

Recovery is especially revealing. What happens after a failed grasp, a blocked path or an unexpected object? A robot that stops and requests help can be easier to live with than one that repeatedly tries the wrong action. Look for evidence of these ordinary interruptions. A claim about future software should remain a future claim until there is a documented release and a demonstration of the relevant behavior.

Use a short ownership interview

For your shortlist, write one sentence for each robot: it performs this task, in this environment, using these computers, with this kind of human involvement. Add a source and checked date. If a manufacturer has not documented an answer, leave it unknown. Filling gaps with assumptions makes a table look complete while making the buying decision worse.

Ask the seller to explain a normal session from power-on to completion, including setup and recovery. Request current documentation for the exact edition. Decide whether the resulting level of independence serves your purpose: entertainment, learning, research or practical assistance. You can then appreciate impressive engineering while still judging the product by the work you expect it to do.