Are companion robots actually autonomous? What they can and can’t do

Task vs. engagement valueThe Ladder measures task completion, not sociability

The Autonomy Ladder measures a robot’s ability to complete a defined task end-to-end within its operating design domain. Level I requires a human to initiate and monitor each operation. Level II adds some sensor-driven adaptation but still requires human framing of the task. Level III requires the robot to complete the task without intervention once the domain is established.

This is a sensible framework for robots whose value is delivered in task completion: a clean floor, a mowed lawn, a scrubbed pool surface. The cleaner the run, the more autonomous the robot, the higher the classification. The measurement and the value are aligned.

Consider the contrast directly. A Level III robot vacuum has a completion signal the owner can verify: the floor is clean, the robot is docked, the app logs a finished run. Every element of the classification points at that observable endpoint. A companion robot has no equivalent signal.

There is no moment at which the ElliQ 3 has finished keeping someone company, no dock-and-done state for Aibo’s relationship with a household. Value accrues in the space between interactions, across weeks, in whether the owner reaches for the robot unprompted. A framework built to reward clean task closure has nothing to measure in a system designed never to close.

Companion robots are not task robots. The ElliQ 3, designed by Intuition Robotics for older adults aging at home, measures its own success by whether a user initiates a conversation, responds to a wellness check, or reports feeling less isolated. The robot cannot ‘complete’ connection.

It can only sustain conditions that make connection more likely. That is not a task with an endpoint; it is a continuous behavioral process. Classifying ElliQ at Level I reflects the Ladder’s assessment that a human must initiate each meaningful interaction. That is accurate. It is also a description of what engagement-based design requires, not a failure.

Robovations classification framework

The Autonomy Ladder: built for task completion

What Level II actually describes on AiboAssisted Autonomy applied to a robot with no task

The Sony Aibo ERS-1000 navigates a room, tracks faces, responds to touch and voice, and builds behavioral patterns over time through Sony’s cloud AI. Owner reports describe the robot developing distinct habits: gravitating toward certain family members, showing reluctance at certain times of day, expressing what Sony calls ’emotion’ via LED eye animations and body posture. The system is not scripted. The behavior emerges from a reinforcement model trained against interaction signals.

None of that maps onto a task. Aibo does not have a floor to clean or a lawn to mow. It has interactions to generate and sustain. Classifying it at Level II reflects the Ladder’s observation that a human must engage for the robot to respond, that Aibo does not initiate complex new behaviors without input, and that there is no ‘task domain’ within which it could operate autonomously.

All of that is correct. What it does not capture is that Aibo’s architecture for generating engagement over a multi-year ownership period is more sophisticated than several Level III task robots that successfully navigate a kitchen floor.

The placement is not wrong. The question is whether the metric is doing useful work for this category.

temi and the hybrid casetemi is a task robot as well as a companion

The temi V3 complicates the pattern in a useful way. temi is a mobile telepresence and ambient computing platform that navigates autonomously to named locations, follows users through a home or facility, and hosts video calls via its built-in display. It classifies at Level II. That placement reflects that temi requires human initiation for most high-value actions, including call initiation and destination assignment.

temi, however, does complete subtasks autonomously: it navigates to a room without guidance, avoids obstacles in real time, and returns to dock without instruction. In a task-only frame, some of these behaviors would approach the threshold for Level III classification.

The platform sits at Level II because those autonomous subtasks exist in service of a human-to-human interaction, not a machine-completion goal. The ‘task’ is the video call; the autonomous navigation is infrastructure for the call, not the deliverable.

This illustrates a broader structural point: companion robots can have sophisticated autonomous subsystems while remaining correctly classified at lower Ladder levels. The Ladder classifies the full use-case, not the most capable subsystem. A robot that navigates autonomously but whose value is delivered entirely by a human caller on the other end of a video feed is not delivering autonomous value, even if its navigation stack is Level III-grade.

Two more robots in the table extend the lesson from different directions. The Amazon Astro pairs genuine autonomous navigation with a home-monitoring task, yet the roaming screen and Alexa persona are what owners describe as the reason it stays in a household rather than the patrol function.

The Enabot EBO X blurs the line the others hold: it maps the home with V-SLAM and patrols autonomously, yet the remote video check-in owners buy it for is driven by a person on the far end of the feed. Its Level II reflects that split between autonomous patrol and human-directed presence, not a single deciding constraint. Same ceiling, different reasons to sit beneath it.

Classification accuracy vs. classification utilityAccurate but not useful is still a problem

The case being made here is not that the Ladder is wrong. The Level I placement for ElliQ and the Level II placements for Aibo, temi, and Amazon Astro are accurate readings of the framework. The case is narrower: for companion and social robots, the classification is accurate and carries almost no information about the robot’s value to its owner.

A consumer shopping for a robot vacuum learns something real from a Level III classification. It means the device will complete a full floor run, dock, and charge without intervention. The classification maps to a concrete ownership experience. A consumer considering the Aibo ERS-1000 at $2,899 learns from the Level II classification that Aibo adapts to its environment and responds to stimuli, but requires human engagement to generate its characteristic behaviors.

That is accurate. It does not tell the consumer whether the robot will feel alive in the room after two years, whether it sustains interaction across a household with multiple people, or whether its emotional model deepens or plateaus over time. Those are the questions that determine whether the purchase delivers on its premise. The Ladder does not reach them.

The collapse is starkest when two companion robots share a rung. The temi V3 and the Aibo ERS-1000 both classify at Level II, yet almost nothing about their ownership overlaps. One is a rolling video-call platform a family buys to check on an aging parent; the other is a robotic dog bought for daily companionship.

A shared Level II tells a prospective buyer they are equivalent in autonomy, which is true, and equivalent as purchases, which is absurd. This is the information gap in miniature: the classification ranks them identically because it can only read the one axis they happen to share, and stays blind to every axis on which they actually differ.

This is a classification scope problem, not a methodology failure. The Autonomy Ladder was designed around task robots. Companion robots were a peripheral category when the Ladder’s core logic was developed. As that category grows and as companion-robot price points rise, the information gap becomes more visible.

[rv_ans_distribution robot=”sony-aibo-ers-1000″ metric=”price,suction,runtime,cost5″]

What useful companion classification requiresClosing the gap would not replace the Ladder

Closing the gap would not mean replacing the Ladder for companion robots. A companion robot’s ability to initiate and sustain behavior without human intervention is still meaningful information.

The ElliQ 3 does proactively prompt wellness check-ins on a schedule; that is a form of autonomous initiation the Level I placement does not fully represent. The Aibo ERS-1000 develops behavioral patterns without explicit instruction; that is more than a pure Level II stimulus-response system.

What companion robots need, alongside the Ladder placement, is a secondary dimension describing engagement architecture: does the behavioral model deepen over time or stay static? Does the robot initiate unprompted behaviors, and how often? How does the system behave after six months of use relative to day one?

These are not autonomy questions. They are persistence and adaptation questions that only become visible in longitudinal owner data. Intuition Robotics has reported that ElliQ engagement tends to persist rather than taper after the initial novelty period, a pattern that no task-completion metric would surface.

The Robovations database currently surfaces Ladder level, readiness status, and score alongside each robot’s classification. For companion robots, those fields do their job accurately. An additional classification dimension specific to engagement persistence would do more useful work for this category specifically than any adjustment to the Ladder itself.

That is a scoped addition, not a framework revision. The Ladder stays exactly as it is for the vacuums, mowers, and pool robots it was built to measure, and companion robots simply gain a second axis that a task ladder was never designed to carry.

The Autonomy Ladder classifies companion robots accurately and incompletely: the Level I and II placements are correct reads of task-completion autonomy, and the category’s real design challenge, sustaining engagement over months and years, sits outside the measurement frame.

[rv_ans_timeline robot=”sony-aibo-ers-1000″]