DFKI-LT - Offline and active gradient-based learning strategies in a pushing scenario

Sergio Roa, Geert-Jan Kruijff
Offline and active gradient-based learning strategies in a pushing scenario
2 Proceedings of the 3rd International Workshop on Evolutionary and Reinforcement Learning for Autonomous Robot Systems, Lisbon, Portugal, o.A., 2010
When operating in the real world, a robot needs to accurately predict the consequences of its own actions. This is important to guide its own behavior, and in adapting it based on feedback from the environment. The paper focuses on a specific problem in this context, namely predicting affordances of simple geometrical objects called polyflaps. A machine learning approach is presented for acquiring models of object movement, resulting from a robot performing pushing actions on a polyflap. Long Short-Term Memory machines (LSTMs) are used to deal with the inherent spatiotemporal nature of this problem. An LSTM is a gradient-based model of a Recurrent Neural Network, and can successively predict a sequence of feature vectors. The paper discusses offline experiments to test the ability of LSTMs to solve the prediction problem considered here. Cross-validation methods are applied as a measure of convergence performance. An active learning method based on Intelligent Adaptive Curiosity is also applied for improving the learning performance of learners trained offline, generating a combination of learners specialized in different sensorimotor spaces after the knowledge transfer.
Files: BibTeX, RoaKruijffERLARS10.pdf