By Piaget's account, the sensorimotor stage in human infants lasts roughly twenty four months [84][83]. In the first four months, reflexive responses begin to organize into coherent motor strategies, sensory modalities are coordinated and attentional mechanisms begin to emerge. Native reflexive responses like the primary walking reflex and the palmar grasp reflex [100] provide primitive sensorimotor behavior that manage appropriate musculo-skeletal structures to do relevant, sensory-driven work in the world. Bruner [102] refers to these types of behaviors as ``preadaptation'' primitives for learning skillful motor policies.
Policies for coordinating multiple sensory and motor modalities appear next in Piaget's account. Primary circular reactions form, which can be viewed as a primitive form of an identity judgment. From four to six months, these primary circular reactions are repeated (assimilation) and perturbed (accommodation) until the infant finds it possible to prolong certain interactions with the world. Between six and eighteen months Piaget believed that these primary circular reactions direct subsequent exploration and influence evolving behavioral models of the world. During this same period, figurative schemas appear to encode information regarding ``types'' or ``classes'' of behavioral relationships in a manner that depends less on specific resources.
Complex robot systems also require some preadaptive structure to organize the acquisition of sensorimotor behavior. Our model (introduced in Section 1.2) describes a mechanism for constructing behavior as sequences of control situations. Control situations, in turn, are combinations of controllers and system resources. Although the details differ significantly across human and robot subjects, we claim that both systems must initially acquire a policy for engaging resources in a coordinated fashion, and that these policies form the most basic model of the agent/world interaction.
The central objective of the research is to test the hypothesis that many situations can be expressed in terms of relatively few, schematic structures that are grounded in physical activity, which we call physical schemas [52][74][73][68][67][66] and the related hypothesis that physical schemas are a sensorimotor foundation for conceptual knowledge. One implication of our theory is that once stable relationships in the world are discovered and captured as physical schemas, an agent may notice other instances of these relationships involving different entities. This is one form of metaphorical extension [66], whereby physical schemas are extended to other physical, and sometimes nonphysical, phenomena. For example, underlying the word ``contain'' is a physical schema of the activities and relationships we call CONTAINMENT. Explicit in this schema is a stable pattern of dynamic activity in which one entity is constrained by another. This, we argue, is learned by direct physical interaction with the world when, for instance, we grasp an object. The entailments of CONTAINMENT are preserved if we consider that water is contained in a cup - it is only the participants in the relationship that have changed. Physical schemas, from this perspective, are a foundation for an enormous range of plausible physical and mental phenomena. The idea of physical schemas is not new [68][67][66], nor is its role in theories of conceptual development [52][74][73], but we are proposing a computational account of physical schemas implemented on robotic platforms.