A common paradigm in artificial intelligence takes the position that intelligent behavior is derived from lots of domain-specific knowledge. However, our efforts at engineering knowledge work only in narrow contexts and produce, therefore, very special purpose artifacts. The tradition in the behavioral sciences, in contrast, appreciates the interactive aspects of knowledge and intelligence, but provide relatively little insight into mechanisms. Recent paradigms emerging in Philosophy, Psychology, Linguistics and Robotics offer a plausible framework for modeling the process of sensorimotor development. Such a framework can function both as a descriptive account of human sensorimotor development and as a design principle for intelligent machines. This proposal advances a computational model for sensorimotor development and explores the implications of such a model. This framework leads to mechanisms for organizing experience as behavioral policies, and a means of segmenting sensory feedback gathered over a sequence of motor activities into ``objects.'' These data structures form the basis, we argue, of a conceptual structure that influences all subsequent cognitive development. This proposal gathers insights from developmental psychology, control theory, reinforcement learning, robotics and artificial intelligence to look for the mechanisms responsible for our very earliest knowledge structures.