Next Step of the "Assembly Plan from Observation" System
A novel robot-programming paradigm called gLearning by Demonstrationh has been proposed. It can automatically generate robot-programs by understanding human demonstration. Users can therefore easily let a robot execute various desired tasks. We propose an "Assembly Plan from Observation" (APO) System, which has the ability to make an assembly task program automatically from a human demonstration. However, it can make this program only because it understands the task object using contact relations. Our next challenge is to understand various tasks by improving the APO system.
First, we focus on the manipulation control model, position-force hybrid control, which applies position control to DOFs not to be constraint to a contact, and applies force control to DOFS to be constraint. Considering that position-force hybrid control is very effective for various tasks, the system should have the ability to divide DOFs into position or force control DOFs. Usually data obtained from a robot vision system includes some errors, and the system needs to correct errors before understanding the task. However, the data is also perturbed by error correction. We propose a method to correct errors without perturbation and to divide DOFs. Using such corrected data, we try to explore a method to analyze divided constraint-free motion.
