Abstract not found

We have been exploring an approach to robot learning based on a hierarchy of types of knowledge of the robot's senses, actions, and spatial environment. This approach grew out of a computational model of the human cognitive map that exploited the distinction between procedural, topological, and metrical knowledge of large-scale space. More recently, the semantic hierarchy approach has been extended to continuous sensorimotor interaction with a continuous environment, demonstrating the fundamental role of identification of distinctive places in robot spatial learning. In this paper, we describe three directions of current research. First, we are scaling up our exploration and map-learning methods from simulated to physical robots. Second, we are developing methods for a tabula rasa robot to explore and learn the properties of an initially uninterpreted sensorimotor system to the point where it can reach the control level of the spatial semantic hierarchy, and hence build a cognitive map...

The Spatial Semantic Hierarchy and its predecessor the TOUR model are theories of robot and human commonsense knowledge of large-scale space: the cognitive map. The focus of these theories is on how spatial knowledge is acquired from experience in the environment, and how it can be used effectively in spite of being incomplete and sometimes incorrect. This essay is a personal reflection on the evolution of these ideas since their beginning early in 1973 while I was a graduate student at the MIT AI Lab. I attempt to describe how, and due to what influences, my understanding of commonsense knowledge of space has changed over the years since then. 1 Prehistory I entered MIT intending to study pure mathematics. I was generally steeped in the ideology of pure mathematics, and I had every intention of staying completely away from practical applications in favor of abstract beauty and elegance. However, on a whim, in Spring of 1973 I took Minsky and Papert’s graduate introduction to Artificial Intelligence. I was immediately hooked. I had always been fascinated by the idea

as a testbed for designing intelligent agents. The system consists of an overall agent architecture and five components within the architecture. The five components are: 1. Goal-Directed Learning (GDL), a decision-theoretic method for selecting learning goals. 2. Probabilistic Bias Evaluation (PBE), a technique for using probabilistic background knowledge to select learning biases for the learning goals. 3. Uniquely Predictive Theories (UPTs) and Probability Computation using Independence (PCI), a probabilistic representation and Bayesian inference method for the agent's theories. 4. A probabilistic learning component, consisting of a heuristic search algorithm and a Bayesian method for evaluating proposed theories. 5. A decision-theoretic probabilistic planner, which searches through the probability space defined by the agent's current theory to select the best action. PAGODA is given as input an initial planning goal (its ove