Toward Learning Robots

Toward Learning Robots

The contributions in Toward Learning Robots address the question of how a robot can be designed to acquire autonomously whatever it needs to realize adequate behavior in a complex environment. In-depth discussions of issues, techniques, and experiments in machine learning focus on improving ease of programming and enhancing robustness in unpredictable and changing environments, given limitations of time and resources available to researchers. The authors show practical progress toward a useful set of abstractions and techniques to describe and automate various aspects of learning in autonomous systems. The close interaction of such a system with the world reveals opportunities for new architectures and learning scenarios and for grounding symbolic representations, though such thorny problems as noise, choice of language, abstraction level of representation, and operationality have to be faced head-on. Contents Introduction: Toward Learning Robots * Learning Reliable Manipulation Strategies without Initial Physical Models * Learning by an Autonomous Agent in the Pushing Domain * A Cost-Sensitive Machine Learning Method for the Approach and Recognize Task * A Robot Exploration and Mapping Strategy Based on a Semantic Hierarchy of Spatial Representations * Understanding Object Motion: Recognition, Learning and Spatiotemporal Reasoning * Learning How to Plan * Robo-Soar: An Integration of External Interaction, Planning, and Learning Using Soar * Foundations of Learning in Autonomous Agents * Prior Knowledge and Autonomous Learning

Towards Real Learning Robots

Reinforcement learning, in a nutshell, is a form of learning that enables the robot to construct a control law by a system of feedback signals that reinforce «electrical path ways» that produce correct response, and conversely wipe-out connections that produce errors. Unfortunately, without biasing, it is a weak learning that presents unreasonable difficulty, especially when it is applied to real robots. The subject of this thesis is to study, for a particular class of problems, the effects of different form of biases on the speed of learning as well as on the quality of final learned policy, and to realize this learning paradigm on a physical robot by appropriately biasing the robot with domain knowledge that determines how much the robot knows about the different parts of its world.

Toward Learning Robots