Companion robots and the autonomy classification problem when the task is interaction

The Robovations Autonomy Ladder grades robots on whether they complete their operating task within a defined domain without human instruction during execution. Level I robots execute scripted routines triggered by explicit human input. Level II robots sustain goal-directed operation within a bounded environment, adapting to changing conditions, but rely on human setup or periodic human confirmation. Level III robots complete end-to-end task cycles without intervention within the operating domain.

For a floor vacuum, this framework maps cleanly onto physical behavior. The robot either navigates the floor autonomously or it does not. For a companion robot, the operating task is social presence, engagement, or monitoring. The question becomes: does the robot sustain that task within a defined interaction domain without continuous human instruction?

ElliQ 3 is the clearest example of how the boundary resolves. Manufactured by Intuition Robotics, ElliQ is purpose-built for older adults as a proactive engagement platform. It initiates check-ins, prompts physical activity, and surfaces health reminders without user input. But its behavioral repertoire is built on a scripted dialogue tree and cloud-hosted response logic. The robot does not generate novel conversational strategies; it selects from authored interaction modules. That is Level I operation: the task is interaction, but the execution is scripted trigger-and-response. The $249 hardware MSRP is paired with a subscription that sustains the content pipeline, and ElliQ 3 carries a Level I classification at Robovations as a result.

Amazon Astro and temi V3 carry Level II classifications because they add a spatial dimension to social presence. Both robots navigate a home or office layout autonomously, tracking individuals across rooms, delivering video call interfaces to wherever a user is located, and initiating check-in routines on a set schedule. The autonomous navigation raises the classification because it expands the operating design domain beyond a fixed docking station and because the robot sustains goal-directed behavior (find the person, initiate contact) without per-instance human instruction.

Astro’s classification at Level II reflects Amazon’s stated design: the robot is intended to function as a home presence and security monitoring device, autonomously patrolling mapped areas and surfacing Alexa functions wherever the user is. Amazon’s product documentation describes autonomous home patrol and self-initiated check-in sequences. The $1,599 MSRP positions it at the lower end of mobile social robots.

Sony Aibo ERS-1000 also lands at Level II. The Aibo’s quadruped locomotion is genuinely autonomous: the robot navigates household spaces, seeks out its owner, responds to voice and touch, and expresses states through physical posture and eye animation. Sony’s product documentation describes autonomous behavior that includes self-initiated play sequences and charging-dock return without instruction. The $2,899 MSRP reflects the biomechanical complexity. But the Aibo is not goal-directed in the functional sense the Ladder rewards; it does not complete a defined task end-to-end. It sustains presence. That is Level II.

Level III requires end-to-end task completion within the operating design domain without human intervention. For a companion robot, the equivalent would be: the robot identifies that a person needs engagement, selects an appropriate interaction strategy, executes it through to resolution, and adapts when the person’s state changes, without any scripted trigger or human confirmation. No consumer companion robot in production achieves this.

The limitation is not computational power or connectivity. It is domain definition. Social interaction does not have a bounded operating design domain in the way a floor map does. The robot cannot know when a task is complete because emotional presence does not have a verifiable terminal state. A floor vacuum can confirm room coverage. A companion robot cannot confirm that the person feels less lonely.

This is not a criticism of the category. It is a structural description of what companion robots are doing and what the Autonomy Ladder is measuring. The robots in this category are genuinely useful within their operating envelopes. ElliQ 3 has documented outcomes in owner reports describing sustained engagement with older adults over multi-month deployments. temi V3 is deployed in clinical and hospitality settings where autonomous navigation extends human reach across physical space. These are real capabilities. They are Level I and Level II capabilities.

ElliQ 3 surfaces a secondary classification issue that temi V3 and Astro avoid: the robot’s meaningful behavior is gated behind a subscription. Without active subscription, the ElliQ hardware provides limited functionality. The interaction pipeline, the health reminders, the engagement content, these are cloud-delivered services, not on-device capabilities. This matters for classification because the Autonomy Ladder grades the robot’s independent capability, and a robot that depends on an active paid cloud service for its core function is not fully autonomous in any sense the Ladder rewards.

This is the same pattern documented in Robovations’ earlier analysis of subscription gating across consumer robot categories. The distinction for companion robots is that the gating is not incremental (as when a vacuum adds a premium map feature behind a paywall) but foundational. The task itself is cloud-delivered. That does not change the Level I classification; it reinforces it.

temi V3 avoids this structure. The robot’s navigation and video-calling functions operate on local network infrastructure and do not require an ongoing subscription to the manufacturer. Amazon Astro’s model is layered: core navigation and presence-monitoring functions operate locally, while Alexa integration depends on Amazon’s cloud services. Both the temi V3 and Astro classifications treat local operational capability as the baseline.