Kinematic structures are currently one of the most important prerequisites for advanced robot technology.
Grasping objects, painting cars or performing household tasks are just a few examples from the manifold space of possible applications.
To execute such operations, the structure of an object has to be knownapriori.In dynamical changing environments, however, this prior specification becomes unfeasible. Therefore, an automated process is required that can reliably extract kinematic structures from visual sensory input.
This work contributes to this effort by acquiring joint types from moving 3d point clouds. A cloud consists hereby of presegmented visual features from object parts or the environment.
The algorithm tries to classify each relationship in the cloud according to different categories of joint types.
Several real world experiments with ordinary objects like doors, drawers, tricycles, and laptops were performed, to test the stability of this approach.