For care robots to be effective, they must interact meaningfully with individuals. Known as human-robot interaction, this involves understanding and responding to user needs. Our goal was to improve this interaction by enabling the robot to reason (‘what-if’, ‘why not?’) using natural language generation underpinned by counterfactual reasoning. This approach helps the robot consider alternate scenarios, fostering safer, more personalised assistance.

What did we do?

During seven co-design workshops, we worked with 25 participants (older adults and caregivers) from both urban and remote locations, including island communities. Using a simulator that replicated a robot interacting with a person, we co-created tasks and iterated the robot’s language. Adopting counterfactual reasoning allowed the robot to reflect on risky or unsuccessful situations and illustrate how altering specific factors could lead to different, “better” outcomes. 

Outcomes and Findings

Building on the simulator framework, we integrated an NLP component with natural language understanding, generation, and counterfactual reasoning. We defined responses, requirements, and intents to enable contextually appropriate conversations. In a final online workshop, participants offered feedback and proposed new tasks, directly shaping our analysis. From these insights, we identified four key considerations:

• Multi-Layered Self Awareness: The robot must account for its own abilities, the user’s physical and emotional state, and the environment, enabling safe, context-sensitive decisions (e.g., postponing a walk if the user is fatigued).

• Dynamic Consent and Evolving Permissions: Seeking explicit permission for unexpected tasks (like cleaning a spill) respects boundaries and adapts to shifting user needs.

• Balance of Safety and Independence: The robot should assist without undermining autonomy, allowing older adults to stay active rather than relying on automatic task completion (but explaining this to the user).

• Feedback Integration: Continual improvement relies on immediate and retrospective feedback, ensuring the robot refines its performance to safeguard user well-being.

Significance

In summary, our exploration of care robots highlights the potential for AI-driven assistance to supplement traditional care and promote independence among older adults. By involving stakeholders in co-design and leveraging counterfactual reasoning, we have identified useful areas of focus for future development and research directions.

This project successfully developed a multi-agent conversational framework designed to enhance human-robot interactions in elder care settings. Through seven co-design workshops with older adults and care workers in urban, rural, and respite care environments, we gathered valuable insights into user needs, leading to key refinements in our approach. The main findings from the workshop were around language/vocabulary, personalisation, data protection and safety. 

For robot simulation, we integrated our multi-agent framework into Legent, an open source robot simulation platform that enables users to interact with the robot via a chat interface. We replaced its existing NLP module with our framework, refining it iteratively through workshop feedback. The agent is capable of tasks such as identifying hazards, resolving identity ambiguities, and requesting clarifications from users when necessary. Additionally, we introduced a dashboard for managing patient personas, allowing the robot to tailor conversations. For instance, if a user is colourblind, the robot adjusts its explanations, using size or surrounding objects rather than colour to describe items.

Our framework, built using OpenAI services and Langchain, features a routing manager based on prototypical networks that dynamically assigns user queries to the most relevant task agent. These advancements contribute to making conversational AI more adaptive, transparent, and user-centered, enhancing its effectiveness in elder care settings.

Co-Design Activities – We conducted 7 workshops engaging elder care professionals and older adults to gather insights and refine our approach. A detailed description of these workshops can be found in the Stakeholder Engagement Section of the report.

Robot-Simulation Environment – We developed a conversational AI framework for elder care robotics, now integrated with the Legent AI Simulator(www.docs.legent.ai) for interaction. We plan to release the codebase soon to support further development and research.

Journal Paper – We are preparing a publication for the JMIR (Journal of Medical Internet Research), which will document findings from our seven workshops and provide implementation details of our conversational AI framework.These deliverables contribute to advancing conversational AI for elderly care, ensuring practical, ethical, and user-centered development.