Learning via imitation and kinematic analogies

Due to the growing complexity of robotic systems and applications,
defining poses and motions for robots is an increasingly difficult problem.
Hand-coding approaches do not provide the required scalability, while
machine learning algorithms based on teacher imitations often lack
generality due to required a priori kinematic transformations. This work
presents an general approach to learn kinematic pose information from a
single point cloud image of teacher with an arbitrary kinematic structure.
For the midterm milestone, the software infrastructure for visualization and
kinematic representation is developed and initial testing with optimization
via evolutionary algorithms are in place.

Category:

People & Blogs

Tags:

• Ly