DFG KE 1686/2-1: Relational exploration, learning and inference - Foundations of autonomous learning in natural environments




Learning and exploration are among the most interesting aspects of intelligent behavior in humans and animals. In this project we address exploration and learning in the context of object manipulation in natural environments—such as a household, offices or factories—where the state can be described in terms of (continuous or discrete) relations between objects. Handling this setting is crucial to advance the state-of-the-art in intelligent systems towards real-world applicability. Existing machine learning and robotics methods, however, largely fail as the inherent world structure in the exponential state spaces is not exploited, while AI approaches typically neglect learning and exploration under uncertain experiences. The core approach of the project is to organize exploration, learning and inference on appropriate relational representations implying strong prior assumptions on the world structure. On these representations we can learn from uncertain experience compact models of action effects that generalize across objects. We transfer existing exploration theories to relational representations—leading to a novel level of explorative behavior that decidedly aims to explore objects to which the current knowledge does not generalize. This project is to our knowledge the first to combine statistical relational learning methods to tackle core problems in intelligent robotics, fueling the hope for a major advance in the field. We will demonstrate our methods on real-world robot platforms manipulating their environments.



Staff Members:

Hadiji, Fabian
Kersting, Kristian
Mladenov, Martin