Bootstrap Learning and Visual Processing Management on Mobile Robots

A central goal of robotics and AI is to enable a team of robots to operate autonomously in the real world and collaborate withhumans over an extended period of time. Though developments in sensor technology have resulted in the deployment of robotsin specific applications the ability to accurately sense and inter act with the environment is still missing. Key challenges to thewidespread deployment of robots include the ability to learn models of environmental features based on sensory inputs, bootstrapoff of the learned models to detect and adapt to environmental changes, and autonomously tailor the sensory processing tothe task at hand. This paper summarizes a comprehensive e ffort towards such bootstrap learning, adaptation, and processingmanagement using visual input. We describe probabilistic algorithms that enable a mobile robot to autonomously plan its actionsto learn models of color distributions and illuminations. The lear ned models are used to detect and adapt to illumination changes.Furthermore, we describe a probabilistic sequential decision-making approach that autonomously tailors the visual processing tothe task at hand. All algorithms are fully implemented and tested on robot platforms in dynamic environments.