Neural Networks for Robotics

Research into bio-inspired, smart multirobot controllers for exploration.

Evolvable neural networks are simplified models of the nervous system inspired by biological evolution and is form of machine learning offering great potential in artificial intelligence applications and “Big Data”. SpaceTREx has established a cutting-edge research program focused on the development of Artificial Neural Networks (ANN) with particular focus on the exploration of space and extreme environments, but with broad applicability to terrestrial environments including controllers for multi-agent systems software and pattern recognition within sensor networks. This venture combines research on Evolutionary Algorithms (EA) and ANNs with High-Performance Computing (HPC). Current research is focused on the development and implementation of new methods for the optimization of network weights/topologies, enabling the discovery of novel solutions to complex tasks through evolving Artificial Neural Tissues (ANT) [Thangavelautham & D’Eleuterio, 2005]. This method has proven superior to standard neural network controllers and human devised controllers applied to multi-robot excavation tasks. Through the application of level-set methods, a sequence of global cost functions are minimized, leading to efficient solution convergence in problems which are intractable using contemporary methods. ANT enables efficient cooperation within multi-agent networks while maintaining robust fault tolerance and graceful degradation. The ultimate goal of this project is to further develop this level-set methodology toward human-competitive capabilities, thus improving the efficiency of space exploration tasks such as resource gathering and data acquisition. The application of such controllers to networks of robots/sensors will lead to great advances in the solution of complex problems such as image classification, tracking, navigation, resource gathering, excavation, and large-scale data analytics.