Pushing cylinders: Expanding on object-based manipulation

Competent assistant robots need to understand their environment to be able to perform skilled manipulation tasks. One way of achieving it is to have knowledge about the physical behaviour of objects and how they respond to certain stimuli. The proposed approach is to have a library of compact predictor-controller pairs for different simple tasks that chained together enable the robot to execute complex actions. In this paper we expand the previous work by adding the compact models for pushing a cylinder over a table to a goal pose. The model was tested and compared against our previous box model with a real humanoid robot by pushing two cylinder shaped and a box shaped object from random starting positions to predefined goals. The robot successfully pushed the objects towards the goals. With this, as the main contribution fo this paper, we have successfully expanded the systems presented on previous works with new capabilities — namely the ability to work with cylinder shaped objects.