A human infant facing the blooming, buzzing confusion of the senses grows up to be an adult with common-sense knowledge of geometry; this knowledge then allows her to describe the shapes of objects, the layouts of places, and the relative locations of things naturally and effortlessly. In robotics, such knowledge is usually built in by a human designer who needs to solve complex engineering problems of sensor calibration and inference. In contrast, this dissertation presents a model for how autonomous agents can form an understanding of geometry the same way infants do: by learning from early unstructured sensorimotor experience. Through a framework called sensorimotor embedding, an agent reconstructs knowledge of its own sensor structure, the local geometry of the world, and the pose of objects within the world. The validity of this knowledge is demonstrated directly through Procrustes analysis and indirectly by using it to solve the mountain car task with different morphologies. The dissertation demonstrates how sensorimotor embedding can serve as a robust approach for acquiring geometric knowledge.

PDF

PhD Thesis, Department of Computer Sciences, The University of Texas at Austin.

@phdthesis{stober:thesis, title={Sensorimotor Embedding: A Developmental Approach to Learning Geometry}, author={Jeremy Stober}, month={May}, school={Department of Computer Sciences, The University of Texas at Austin}, address={Austin, TX}, url="http://www.cs.utexas.edu/users/ai-lab?stober:thesis", year={2015} }