Bringing up robot: F undamental mechanisms forcreating a self-motiv ated, self-organizing architecture In this paper we describe an intrinsic developmental algorithm designed to allow a mobile robot to incrementally progress through levels of increasingly sophisticated behavior. We believe that the core ingredients for such a developmental algorithm are abstractions, anticipations, and self-motivations. We propose a multi-level, cascaded discovery and control architecture that includes these core ingredients. Toward this proposal we explore two novel models: a governor for automatically regulating the training of a neural network; and a path-planning neural network driven by patterns of mental states which represent protogoals.1 Introduction Most intelligent robot control systems begin with the goal of creating a robot to carry out humanissued tasks. While these tasks v ary in difficulty , they must, by their very nature, involve abstract concepts. For example, typical tasks might be: go to a specific location, identify an object, or pick up an object. Attempting to directly achieve the goal of carrying out human commands creates basic assumptions about the architectural design of a robot. We call this philosophy task-oriented design. Within the task-oriented design paradigm, there are two competing methodologies: top-down, and bottom-up. T op-down designers apply computational algorithms that can be carried out on the robots so as to accomplish a given task. There is a range of computational models employed in robotics: dead reckoning (e.g., using internal measures of space), sensor fusion, behavior fusion, and symbolic logic. Bottom-up designers again usually take the task to be performed by the robot as a pre-specified assumption. However, the control architecture of the robot is designed in a bottom-up fashion.Examples include subsumption architectures, supervised learning schemes, and evolutionary computation. We believe that a significant pitfall exists in both the top-down and bottom-up task-oriented robot design methodologies: inherent anthropomorphic bias. This bias refers to the design of prespecified robot tasks: traditional research in the design of intelligent robots has attempted to get robots to do the tasks a human can, and do it in a human-centered manner. Historically , this methodology started out by imitating the physical actions of a child playing with blocks. A task was decomposed into a planning problem, and then, with a robot equipped with an arm and a gripper, the robot was asked to manipulate specic blocks. The inherent anthropomorphic bias existed by design, since the issue was to explore models of intelligent behavior. The pitfall in this approach is that the symbolic modeling of behavior is based on the capabilities of a human body and human concepts. Both capabilities may be inappropriate assumptions for the physical body and experiences of the robot.Furthermore, even if we could build a robot with a human-like body and senses, it is not clear that we can jump straight to the abstract task at hand. Many control issues need to be solved in order to have a robotic system carry out even the simplest of tasks. After a half-century of continued research, the artificial intelligence and robotics communities are still far from developing any type of general purpose intelligent system. Recently , a new approach called developmental robotics is being applied to the design of robot behaviors. In this approach, an artifact under the control of an intrinsic developmental algorithm discovers capabilities through autonomous real-time interactions with its environment using its own sensors and effectors. That is, given a physical robot or an artifact, behaviors (as well as mental capabilities) are gr own using a developmental algorithm. The kinds of behaviors and mental capabilities exhibited are not explicitly specified. The focus is mainly on the intrinsic developmental algorithm and the computational models that allow an artifact to grow. A developmental approach to robotics is partly an attempt to eliminate anthropomorphic bias.By exploring the nature of development, the robot is essentially freed from the task of achieving a pre-specified goal. As long as the intrinsic developmental algorithm demonstrates growing behavior there is no need to pre-specify any particular task for the robot to perform. Indeed, it is the goal of developmental robotics to explore the range of tasks that can be learned (or grown) by a robot, givena specific developmental algorithm and a control architecture. This paper outlines our approach to a developmental robotics program and two experiments toward an implementation.2 Overview of a Developmental Robotics Paradigm The ultimate goal of our developmental robotics program is to design a control architecture that could be installed within a robot so that when that robot is turned on for the first time, it initia