A Tree-Based World Model for Reducing System Complexity in Autonomous Mobile Manipulation

Abstract

Mobile manipulation tasks in unstructured environments remain challenging for autonomous robots. The need to employ a diverse set of software and hardware components to solve the various subtasks inevitably increases system complexity. Knowledge exchange among such diverse components renders them highly coupled, reduces communication efficiency, and makes the knowledge less accessible. To overcome these challenges, we propose AIMM-WM, a central world model as a single source of truth having an abstracted geometric tree structure. Despite its concise, efficient state representation, AIMM-WM is able to provide a wide range of information from low-level geometries to highly abstracted symbols and is interfaced with diverse components for navigation, motion planning, perception, decision-making, and mission control. We evaluate the performance of AIMM-WM from the real use case of our Lightweight Rover Unit during the four-week Moon-analogue demo mission on Mt. Etna, Italy.