Autonomous Reconfiguration Planning in Modular Robots

Autonomous Reconfiguration Planning in Modular Robots

Abstract

Self-Reconfiguring Robots (SRR) are composed of many modules that have the ability to autonomously attach and detach, enabling adaptation to a variety of tasks in unknown surroundings. In order to change shape into a particular form, an SRR must plan a sequence of module movements. This reconfiguration problem is challenging because of the many mechanical degrees of freedom and the resultant large number of possible SRR configurations which contribute to a vast and high-dimensional search space. To support the operation of an SRR in a practical environment, reconfiguration planners must satisfy a number of properties including decentralized computation, parallel motion of modules, real-time execution and planning in the native kinematic space of the module mechanism. The ultimate solution would be a general planner that solves reconfigurations of arbitrary module designs in this way. This thesis has taken a step in this direction of generality by developing a reconfiguration planner of the 3R module that is easily instantiable to other module types. The planner is scalable and decentralized, and considers the native kinematics of a module. It demonstrates the ability to coordinate the parallel motion of modules and executes in real-time. The thesis presents both centralized and decentralized implementations of the algorithm along with performance evaluation for several reconfiguration examples, as well as an analysis of achievable SRR reconfigurations.

Avatar
Rowan McAllister
Postdoc in Machine Learning

My research interests include autonomous vehicles, reinforcement learning, and probabilistic modelling.