Learning spatial preconditions of manipulation skills using random forests

Robots working in everyday and unstructured environments will need to perform manipulation skills using different sets of objects. To determine if a manipulation skill can be executed in a given situation, a robot will need to learn the preconditions of the skill. The robot will need to check both that the required objects are present in the scene and that they are arranged in a suitable manner for the skill to be executed. We propose a random forest approach to learn the set of spatial configurations of objects that fulfill a skill's preconditions. We also explore how parts of objects and interactions between parts can be incorporated into the scene models to improve the generalization performance. The proposed approach was evaluated on the preconditions of six manipulation skills. The experiments show that using the ensemble approach, and including the parts and interactions, results in an increase in accuracy of 16.4%.