Learning manipulative skills using a POMDP framework

Uncertainty is one of the most difficult factors to handle when we wish to develop an algorithm for robot motion planning in real circumstances. This paper presents a solution for a robot to deal with "lack of observation" in the scope of object manipulation. Considering a robotic bartender that picks up a glass filled with an unknown amount of water and tilts it to pour the water into empty glasses, the question is how to find the angle at which the giver glass is tilted to pour the water to the same level in each of empty receiver glasses. To achieve the objective, the amount of water poured is represented with mathematical models of non-linear functions, and numerical simulations are performed using the point-based value iteration algorithm for POMDP to get corresponding tilting angles of the giver glass. We found that the experimental result accuracy reaches 99.025% of similarity with the assumed mathematical model, given an initial tilting angle and the water level in the initial glass. We further verified the validity of the proposed algorithm through dynamic simulations.