Touch based POMDP manipulation via sequential submodular optimization

Ngo Anh Vien, Marc Toussaint

Abstract: Exploiting the submodularity of entropy-related objectives has recently led to a series of successes in machine learning and sequential decision making. Its generalized framework, adaptive submodularity, has later been introduced to deal with uncertainty and partially observability, achieving near-optimal performance with simple greedy policies. As a consequence, adaptive submodularity is in principle a promising candidate for efficient touch-based localization in robotics. However, applying that method directly on the motion level shows poor scaling with the dimensionality of the system. Being motivated by hierarchical partially observable Markov decision process (POMDP) planning, we integrate an action hierarchy into the existing adaptive submodularity framework. The proposed algorithm is expected to effectively generate uncertainty-reducing actions with the help from an action hierarchy. Experimental results on both, a simulated robot and a Willow Garage PR2 platform, demonstrate the efficiency of our algorithm.