My lab focuses on applying decentralized deep learning techniques to optimize multi-robot perception performance and resiliency. This research can be used as a framework for applying Graph Neural Networks to robotic networks as a way to reduce bandwidth requirements and increase performance.
Especially in the age of AI, understanding and optimizing the way robots navigate through problems with the perception of their environment is extremely important. By harnessing the power of Graph Neural Networks (GNNs), these risks stemmed from sensor noise and lack of information can be mitigated. Unlike more traditional network structures involving convolution, Graph Neural Networks are part of a category called Decentralized neural networks. Instead of neurons linked from layer to layer, GNNs consist of nodes connected by links called edges. Instead of using convolution to aggregate features from neurons, GNNs use a process called message passing that allows nodes to update their representation based on that of nodes surrounding them. This project aims to apply GNNs to the problem of multi-robot perception, where aggregating features from individual nodes (or robots) can prove useful in combating useless data or sensor noise. The first method created in this project involves encoding robot position and orientation into messages, meaning that individual nodes can be aware of which other nodes are physically closer to them, allowing for data from drones in proximity to be shared. The second method involves gauging similarity in features in distinct nodes, allowing for other nodes with alike features to be highly weighted and factor more into the makeup of an individual node. When measured against a baseline method without encoding, the two proposed methods were able to accurately represent the subject using depth charts and semantic segmentation. The results also revealed that GNNs relied on less computing power due to message passing than what would have been required to directly share sensor data, proving that the proposed methods are both effective and efficient. With GNNs being a relatively new area of research, this project has shown that they have the potential to be a powerful tool in the field of AI. Drones employing this method can be more effective in doing their jobs, as well as being more resilient to obstacles that might arise along the way.Â
ISR Poster.mp4Johnson_Charles_poster.pdf