Learning for Robot-centric Autonomy

Abstract

Autonomy is a foundational capability that frees robots from confined workspaces and lets them interact with the open world. The traditional approach to robot autonomy has relied heavily on a world-centric approach: building a global, geometrically accurate map and using it for localization and planning. However, this approach often proves inadequate or impractical in many real-world applications. This thesis adopts a robot-centric perspective to autonomy, addressing the challenges across three distinctive scales: (1) Globally, we learn to compress visual experiences into sparse, topological scene representations for long-horizon navigation; (2) At the semi-local level, we develop perception systems that reason about the traversability of the terrains around the robot to achieve robust off-road navigation; (3) Locally, we learn end-to-end perception-action models to navigate a robot to any object with high precision. We demonstrate the real-time performance of our approaches across diverse robotic platforms, highlighting the applicability and generalizability of these methods.