UC Davis

Nowadays, robots have engaged in our daily lives no matter whether in transportation, manufacturing, military, family, education, or entertainment. Because a variety of robots have been invented in distinct environments to help human beings increase working efficiency, avoid injury in dangerous surroundings, or explore unreachable areas, more and more robots have been employed as a robot team to achieve the same goal instead of using a single robot. The use of multiple heterogeneous robots may enhance efficiency, robustness, flexibility, tolerance, and economic benefits now that each skillful robot may complete individual tasks or accomplish cooperative tasks simultaneously in a mission even though failure happens on certain robots or equipment. Multi-robot control has become much more significant than in the past decades. Although numerous structures to build centralized, decentralized, and distributed multi-robot systems (MRSs) and various approaches to coordination or multi-robot task allocation (MRTA) have been proposed, most of them may be unique in specified applications under assumed conditions. Integrating these techniques into different research areas is usually challenging. Even though certain structures and traditional or modern strategies could be combined to apply to additional applications, several challenges are still NP-hard, for example, dynamic environments, optimal solutions, parameter formalization, and parameterizing of robot abilities. An artificial intelligence (AI) method might solve these problems. With the increasing speed of computer processing, machine-learning approaches have been commonly utilized in robotics while few of them were implemented in MRTA. In this dissertation, a novel approach to MRTA with Bayesian Networks (BNs) and a generic agent-based framework for MRSs are proposed to overcome the aforementioned difficulties. The BN of a task could be easily built, and the conditional probability table (CPT) of a BN could be logically established. The success rate of a task could be calculated by following Bayes’ theorem, and tasks could be suitably allocated depending on the collected success rate. Agents in the generic agent-based framework are categorized into agencies. With this framework, a diversity of MRSs could be constructed, and advanced methodology could be implemented. Low-cost, educational hardware robots are exploited to demonstrate our approaches to MRS construction and MRTA in search-and-rescue missions under dynamic environments. The result of MRTA is acceptable with a reasonable setup before training, and the result of MRTA becomes near-optimal after training with a large data set. In the future, we may apply this new MRTA approach to a static environment, a homogeneous robot team, a large-scale robot team, and other types of MRTA problems to analyze extensibility. Researchers may include their state-of-the-art methods in the generic framework and combine our MRTA approach in miscellaneous applications, such as patrolling, surveillance, cleaning, warehouse systems, manufacturing, and exploration to examine flexibility.