Autonomous Robot Learning of Foundational Representations Benjamin Kuipers with Jefferson Provost and Joseph Modayil How does a baby human or robot get knowledge of its own The baby assailed by eyes ears nose skin and entrails at once feels it all as one great blooming buzzing confusion William James 1890 University of Texas at Austin 27 August 2007 Developmental Robotics A variety of baby robots have been created to study foundational learning Baby Robots The RobotCub project iCub Brian Scazellatti s Nico at Yale Minoru Asada et al CB2 Osaka U The mechanical engineering is impressive but the real difficulty is the learning Its knowledge representation must be learned not programmed Our Gedankenexperiment Imagine a baby robot a learning agent born with uninterpreted sensors and effectors It has only pixel level experience Disorganized collection of sensor elements Incremental motor signals How does it learn object level concepts Places paths objects actions etc The macro scale components of adult knowledge The Gedankenexperiment 2 In biological reality some knowledge is innate to the individual Innate knowledge is learned by the species over evolutionary time Breadth first search We pretend that it is learned by the individual Depth first search The gedanken experiment helps illuminate the knowledge and how it can be learned 1 The Gedankenexperiment 3 Current research strategy Devise ways to learn foundational concepts using only general purpose statistical learning methods Future Define a toolkit of statistical learning methods Generate all learnable concepts Search for the most productive concepts The Problem of Learning New Ontology An Ontology is the Foundation for Knowledge Representation An ontology specifies the categories that individuals can belong to the sets variables can be quantified over and the relations that can be defined over those categories Axioms embody the content of knowledge The ontology is the language of objects and relations for expressing the axioms Some authors include axioms in the ontology One Way to Abstract Experience Discrete states and places are abstracted from patterns of continuous behavior How does a learning agent robot or baby get from a pixel ontology of low level sensation to an object ontology of high level concepts How can an agent possibly learn to represent new types of things Explaining Sensor Readings Space is the minimal explanation for similarities among sense values Correlations among pixel values gives the structure of sensory arrays How pixels change in response to motor signals gives the structure of the motor system Pierce Kuipers AIJ 1997 See also Philipona et al 2003a b and Olsson Nehaniv Polani 2006 Life Long Learning Concepts of space objects and actions act as a constant foundation for knowledge They are grounded in sensory and motor interaction with the environment Senses and motors change over a lifetime The grounding of these concepts must adapt 2 Lassie sees the world with a Laser Rangefinder Laser Rangefinder Image 180 narrow beams at 1 intervals 180 ranges over 180 planar field of view About 13 above the ground plane 10 12 scans per second Disorganized Sensor 180 Pixels Structured Sensor Array The Egocentric Range Image The World Centered Range Image 3 The World Centered Range Image Occupancy Grid Statistical Learning Methods Used Objects as Explanations Correlation time series and histograms Agglomerative clustering Multidimensional scaling Dimensionality reduction PCA Isomap Sensory flow Image matching ICP Markov localization max likelihood pose Identify Dynamic Sensor Returns A static world explains most observations So focus on the discrepancies Cluster in space Track over time Gather observations to make shape models Modayil Kuipers 2004 2006 2007 Clustering into Objects 4 Track Objects over Time Describe the Scene Describe the scene in terms of Static world Robot s own pose Object in a fixed position Object and trajectory Individual objects Learning Object Shapes Learning Object Categories Clustering shapes by perceptual features Merge range scans to get shape models Cluster shapes to get object categories Learn about Actions Learn actions to affect objects Learn Learn about Actions Learn their properties Learn to plan Qualitative description of effect Bounds on prerequisite state Control law to perform the action For a mobile robot that can move and push Move to desired point in nearby space Turn to face object Push Move to get object to move also 5 Summary Concept Learning Space is learned as a minimal explanation for sensory correlations Objects are learned as a minimal explanation for discrepancies from fixed value model Actions are learned as minimal descriptions of motions interacting with objects Plans combine actions to achieve goals Roadmap Introduction and Overview Learning from uninterpreted experience Pierce Kuipers 1997 Abstraction of views actions and states Provost Kuipers Miikkulainen 2006 Learning objects and actions Modayil Kuipers 2004 2006 2007 6
View Full Document
Unlocking...