Watch this video from Pieter Abbeel’s group in Berkeley.

Given Abbeel’s research curriculum, I jumped to reading the preprint to see if there was some learning involved. Unfortunately, it seems that the techniques used are
fairly standard (precise state estimation + giant state graph + motion planning). Still, it is a very impressive application.

The contribution highlighted by the paper is the detection of appropriate folding points. While the robot is holding and slowly rotating the cloth, several cameras observe the scene and can reconstruct a 3D model. The analysis of the 3D model allows to detect the towel’s borders, which are the appropriate folding points.

The paper certainly shows that Willow Garage’s robots, and accompanying software, are a robust platform for developing complex applications. For example, this application involves interplay between robot locomotion, stereo vision, and motion planning. The robot must move its arms to grab the towel at the desired folding point, but, at the same time, it must not impact it in other parts. This implies that the (a priori unknown) shape of the towel must be included in the motion planner’s world representation.

All in all, it’s one step towards the robotic maid!