POMDP Planning Under Object Composition Uncertainty: Application to Robotic Manipulation

Manipulating unknown objects in a cluttered environment is difficult because segmentation of the scene into objects, that is, object composition, uncertain. Due to uncertainty, prior work has either identified “best” composition and decided on manipulation actions accordingly or tried greedily gather information about composition. We instead, first, use different possible compositions planning, second, utilize provided by robot actions, third, consider effect competing hypotheses desired task. cast planning problem as partially observable Markov decision process (POMDP) plans over hypotheses. The POMDP chooses action maximizes long-term expected task-specific utility, while doing so, considers informative succeeding In simulation physical robotic experiments, probabilistic approach outperforms using most likely long term greedy making.