A Heuristic Algorithm with Falling Prediction Mechanism for Pose Planning in Irregular Objects Oriented Vertical Stacking Task

In the domain of automatic construction either terrestrial or extraterrestrial, the unmanned system is requested to manipulate objects of regular or irregular shape, seen or unseen, in order to achieve various goals, for instance, collecting geological samples, constructing infrastructure of different types for distinct meets. Vertical stacking is a sub-mission among auto construction and emphasizes the understanding of object stability.This paper focuses on the pose planning part of vertical stacking work-flow and proposes three points to improve the state-of-art: (1) proposing a Main Convex-Hull Faces Based initial posture selecting method instead of randomly setting; (2) Adding a forth object-irrelevant stability filtering layer to the original pose filter so that more stable poses are likely to be selected; (3) constructing a falling predictor based on RUSBoost classifier in the stable pose generation so that algorithm efficiency is improved. We execute a series of simulation experiments to validate the effectiveness of our improvements on two sets of irregular object model sets, one with relatively flat parallel surfaces which favours vertical stacking, another with more general shape which deeply increases task difficulty. The results show that our improvements do strengthen stacking capacity especially the generality of objects manipulated and meanwhile enhance the efficiency.