Word embeddings have triggered great advances in natural language processing for non-embodied systems such as scene describers. Embeddings may similarly advance natural language understanding in robots, as long as those robots preserve the semantic structure of an embedding corpus in their actions. That is, a robot must act similarly when it hears ‘jump’ or ‘hop’ and differently when it hears ‘crouch’ or ‘launch’. This could help a robot learn language because it would immediately obey an unknown word such as ‘hop’ if it had been trained to obey ‘jump’. However, ensuring such alignment between semantic and behavioral structure is currently an open problem. In previous work we showed that the choice of a robot's mechanical structure can facilitate or obstruct a machine learning algorithm's ability to induce semantic and behavioral alignment. That work however required the investigator to create a loss function for each natural language command, including those for which formal definitions are elusive, such as ‘be interesting’. A more scalable approach is to bypass loss functions altogether by inviting non-experts to supply their own commands and reward robots that obey them. Here we found that more semantic and behavioral alignment existed among robots reinforced under popular commands than among robots reinforced under less popular commands. This suggests the crowd either chose alignment-inducing commands and/or preferred robots that acted similarly under similar commands. This may pave the way to scalable human-robot interaction by avoiding loss function construction and increasing the probability of zero-shot obedience to previously unheard commands.

Anticipation is a skill that enables complex decision making in humans and other biological agents. We review different implementations of anticipatory behavior in robots and give an overview on anticipation in biological systems. Based on an example of anticipatory behavior in humanoid robots, we discuss decision making and anticipation in artificial agents. We show that anticipation can enable fast decisions in highly dynamic and complex situations. Our findings are supported by experimental results performed in simulation and on real robots in large scale experiments.

We propose to designate as dynamic interactive artificial intelligence (dAI) a cross-section of existing work in artificially designed and artificially evolved systems meant for minimal forms of interaction with human users. This approach borrows principles from artificial life and human movement science to avoid pitfalls of traditional AI. Counter to tradition, it prioritizes user-machine inter-dependence over autonomy. It starts small and relies on incremental growth instead of trying to implement advanced complete functionality. It assumes a perceptual ontology founded on movement coordination rather than object classification. Its development process is better described as reverse self-organization rather than reverse engineering. dAI can be viewed as a precursor to or pre-condition for enactive AI and an alternative to traditional frameworks grounded on information representation. We then give examples from our work in human movement science where we have used minimal dynamic interactive agents to induce specific beneficial effects in human participants’ movement skills. We also show how dAI can be exploited by both connectionist and symbolic AI.

Active inference introduces a theory describing action-perception loops via the minimisation of variational free energy or, under simplifying assumptions, (weighted) prediction error. Recently, active inference has been proposed as part of a new and unifying framework in the cognitive sciences: predictive processing. Predictive processing is often associated with traditional computational theories of the mind, strongly relying on internal representations presented in the form of generative models thought to explain different functions of living and cognitive systems. In this work, we introduce an active inference formulation of the Watt centrifugal governor, a system often portrayed as the canonical “anti-representational” metaphor for cognition. We identify a generative model of a steam engine for the governor, and derive a set of equations describing “perception” and “action” processes as a form of prediction error minimisation. In doing so, we firstly challenge the idea of generative models as explicit internal representations for cognitive systems, suggesting that such models serve only as implicit descriptions for an observer. Secondly, we consider current proposals of predictive processing as a theory of cognition, focusing on some of its potential shortcomings and in particular on the idea that virtually any system admits a description in terms of prediction error minimisation, suggesting that this theory may offer limited explanatory power for cognitive systems. Finally, as a silver lining we emphasise the instrumental role this framework can nonetheless play as a mathematical tool for modelling cognitive architectures interpreted in terms of Bayesian (active) inference.

We assess existing attempts to build emotions and feelings in machines. We review our recently proposed design for machines possessing analogues of biological feeling. Key to our proposal is a homeostatic architecture that regulates internal states to maintain conditions compatible with life. In a first implementation of our design, we present results from a model of synaptic homeostasis in artificial neural networks. We introduce direct consequences to the network's function as a result of its own information processing activity. This model illustrates the benefits that may accrue to a homeostatic learner when it is placed in a needful and vulnerable relation to the objects over which it computes.

Body movement and proprioception are inextricably linked. Movement produces continuous high-dimensional ensembles of afferent information that provide an internal proprioceptive body representation and its relationship to the environment. Motor function is amenable to recording and interpretation and has been relatively well studied. However, we do not yet understand how physiological proprioceptive afferents contribute to internal body representations, neuromuscular control, and even a sense of agency and self. Proprioceptive and motor signals have often been seen as separate, and to be combined mainly to close feedback loops for neuromuscular control. In contrast, ‘active sensing,’ is an emergent concept for dynamically blending sensory and motor signals. We extend and formalize active sensing into an integrative approach—–born out of a neuromechanical perspective—that sees proprioceptive and motor signals as integral parts of the same functional and perceptual continuum we call the Sensory-Motor Gestalt . The Sensory-Motor Gestalt combines formalisms of physics, state estimation, biomechanics, differential geometry, and physiology to understand the emergence of the self in the context of proprioception and motor actions in the physical world. Proprioception, by defining body state , defines feasible (continuous or discrete) motor actions compatible with that state and the environment. Conversely, motor actions produce subsequent, often predictable, body states. This syntactical relationship leads to an epistemological continuum that spans body state, feasible behavior, agency, identity, and sense of self in organisms and robots.