Probabilistic Action/Observation Planning for Playing Yamakuzushi
Tomohiro Motoda, Weiwei Wan, Kensuke Harada · 2020
We propose a method for observation and action planning for manipulating piled objects under uncertainty focusing on a game of extracting one object from the pile with one finger, which we call Yamakuzushi. In cluttered scenes, an object is often occluded by the others, and it is difficult to recognize objects correctly and to manipulate an object as we want. By discretizing the object state, we plan both the observation pose and sliding action of an object based on a partially observable Markov decision process (POMDP), which can handle uncertainty. We experimentally confirmed that a robot plays Yamakuzushi by finding the correct observation pose and correct action to the stacked pieces and by taking one of the pieces out safely.