Date of Award
Doctor of Philosophy (PhD)
Dr. Prithviraj Dasgupta
Modular self-reconfigurable robots (MSRs) are composed of multiple robotic modules which can change their connections with each other to take different shapes, commonly known as configurations. Forming different configurations helps the MSR to accomplish different types of tasks in different environments. In this dissertation, we study three different problems in MSRs: partitioning of modules, configuration formation planning and locomotion learning, and we propose algorithmic solutions to solve these problems.
Partitioning of modules is a decision-making problem for MSRs where each module decides which partition or team of modules it should be in. To find the best set of partitions is a NP-complete problem. We propose game theory based both centralized and distributed solutions to solve this problem. Once the modules know which set of modules they should team-up with, they self-aggregate to form a specific shaped configuration, known as the configuration formation planning problem. Modules can be either singletons or connected in smaller configurations from which they need to form the target configuration. The configuration formation problem is difficult as multiple modules may select the same location in the target configuration to move to which might result in occlusion and consequently failure of the configuration formation process. On the other hand, if the modules are already in connected configurations in the beginning, then it would be beneficial to preserve those initial configurations for placing them into the target configuration as disconnections and re-connections are costly operations. We propose solutions based on an auction-like algorithm and (sub) graph-isomorphism technique to solve the configuration formation problem.
Once the configuration is built, the MSR needs to move towards its goal location as a whole configuration for completing its task. If the configuration’s shape and size is not known a priori, then planning its locomotion is a difficult task as it needs to learn the locomotion pattern in dynamic time – the problem is known as adaptive locomotion learning. We have proposed reinforcement learning based fault-tolerant solutions for locomotion learning by MSRs.
Dutta, Ayan, "Algorithms for Modular Self-reconfigurable Robots: Decision Making, Planning, and Learning" (2017). Student Work. 2913.
Files over 3MB may be slow to open. For best results, right-click and select "save as..."