Building a robotic self

I wake up in the morning, I see my hand and I know it is me. This seemingly simple skill lies at the foundation of our agentive selves. We assume that the development of the self, in both biological and synthetic agents, is grounded in sensorimotor embodiment and inductive learning biases, as well as in the principle of predictive coding. Here, we explore the core inductive biases and process components for building a robotic minimal self. We propose that the formation of a hierarchical, generative model of the self is driven by active inference processes that interactively infer both an endogenous contextual frame and behaviour within this frame. Accordingly, experimental implementations in humanoid robots show how body ownership and agency can develop as foundational computational components. This work on building a robotic minimal self outlines future research directions in mechanistic modelling of the development of an artificial self. To probe the depth of this development, we sketch out embodied Turing tests that challenge complex problem-solving skills, as well as social intelligence and theory of mind. We suggest that uncertainty resolution and the social domain may be promising avenues for further research in the near future. This article is part of the discussion meeting issue ‘From being to identity: analysis and synthesis of the self’.