Description of Research A. Research Objectives

A planetary rover is a semi-autonomous robot that explores an unknown environment, performing experiments and collecting samples to bring back to its base. A difficult design trade-off between autonomy and safety must be solved with very limited energy and computational resources. We propose a principled strategy whereby the rover incrementally learns a map of its environment consisting of a topological network of places and paths among which it can travel safely and autonomously. A mission to another planet is an enormously expensive one-shot scientific experiment, in which the rover has limited resources of energy, computation, and expected lifespan. Autonomy is essential. A rover on Mars faces a round-trip communication delay of 10-40 minutes, making purely teleoperated motion impractical. At the same time, an unknown environment may contain hazards that a robot is unable to perceive, or to recognize, or to cope with. Safety is also essential, and requires human supervision. Substantial research effort has gone into developing system architectures to support increased autonomy by spacecraft and rovers [3, 7, 18, 17, 16]. That work focuses on the ability to diagnose and respond appropriately to unexpected environmental features and system faults, using techniques such as conditional plans. This is an essential tactical capability for successful autonomy. We focus instead on the strategic question of how to represent knowledge of the spatial environment and how to use that knowledge in pursuit of navigation and scientific goals. There has also been valuable work on how to improve the autonomous spatial navigation capabilities of rovers [14, 15]. Implicit in this work, and in some prominent recent work on robot mapping (e.g., [13]), is the assumption that a " map " of an environment is necessarily a single global frame of reference within which the positions of the rover and its targets (waypoints and scientific goals) are represented as precisely as