3D Probabilistic Representations for Vision and Action
Justus Piater, Renaud J. Detry · ORBi (University of Liège) · 2008
Autonomous robots must be able to construct their own representations that enable them to interact successfully with their environment. In less-than-tightly controlled environments, adequate management of (perceptual and action-related) uncertainty is crucial. We present a framework for 3D visual representations that can be learned from visual training data without requiring external supervision. Once obtained, such representations can be used for fundamental interactive tasks such as object detection, recognition, and pose estimation. Moreover, they can serve as a basis for learning manipulative interaction.