Grounding Theory of Mind in Embodied AI

Embodied AI studies agents that learn, perceive, and act through physical interaction with the world. Learning from human demonstrations is especially promising given the abundance of observable human behavior, but transferring knowledge from humans to robots remains challenging due to differences in embodiment, capability, and perspective. Inverse reinforcement learning (IRL) provides a mechanism for inferring goals from behavior, yet it often conflates true intentions with patterns shaped by physical or cognitive constraints. Human actions reflect not only what people want to achieve but also what is feasible or comfortable for them to do. We introduce PG-ToM, a perspective that treats physical and cognitive limitations as causal factors in action generation and integrates this view with Theory of Mind (ToM). Rather than attributing all behavioral regularities to preferences, PG-ToM frames actions as arising from the joint influence of goals and embodiment, allowing a learner to conceptually separate task objectives from feasibility-driven adaptations. We pose two research questions: (1) How does embodiment shape ToM in artificial agents? (2) How can ToM improve embodied AI's learning and adaptation in dynamic, interactive settings? Using a human-robot dishwashing scenario, we discuss how embodiment grounds ToM and how ToM enhances coordination, learning efficiency, and generalization. While the discussion centers on a lowstakes household task, the ideas extend to broader settings where interpreting human behavior in context is critical for effective human-robot interaction.

Keywords: Artificial intelligence, embodied AI, theory of mind.