Decision-making problems in computationally constrained robot perception



Journal Title

Journal ISSN

Volume Title



Despite recent advances in robotic perception, tasks such as object detection continue to remain challenging on edge devices, especially when real-time operation is required, due to limited processing power and a high volume of input data to process. This thesis examines these challenges from the points of view of architectural design and training of neural networks, the software involved in deploying perception models and algorithms, and decision-making. It does so in the domain of robot soccer and in the context of the RoboCup SPL competition, where only limited, mobile scale hardware is available, and real-time perception is necessary. The thesis makes the following contributions: 1) the design of YOLO-based network architectures and training of lightweight, deep object detection models based on the YOLO paradigm for robot soccer, and the realization of a realtime, on-robot vision system incorporating these object detectors. 2) the design and implementation of a software framework for robot perception based on the Robot Operating System (ROS) that is conducive to usage in the RoboCup SPL, and as a research platform for robot perception under computational constraints. 3) a formalization of the problem of perceptual decision-making, wherein an agent needs to makes choices that decide how visual information is processed. 4) Investigative analysis through experiments conducted in the robot soccer domain showing that perceptual decision-making can enhance task performance through better utilization of limited time and computational resources for perception.


LCSH Subject Headings