Kinesthetic Learning Based on Fast Marching Square Method for Manipulation
Applied Science

The advancement of robotics in recent years has driven the growth of
robotic applications for more complex tasks requiring manipulation
capabilities. Recent works have focused on adapting learning methods to
manipulation applications which are stochastic and may not converge. In
this paper, a kinesthetic learning method based on fast marching square
is presented. This method poses great advantages such as ensuring
convergence and is based on learning from the experience of a human
demonstrator. For this purpose, the demonstrator teaches paths by
physically guiding one of the UR3 arms of a mobile manipulator. After
this first phase, the fast marching Learning method is used to make the
robot learn from this experience. As a novelty, an auto-learning
functionality is presented, which provides the kinesthetic learning
algorithm with an exploration capacity. The base of this algorithm is
not only using the information provided by the taught trajectories, but
also expanding its ability in order to explore unknown states of the
environment. The effectiveness of the proposed method has been evaluated
through simulations in 2D and 3D environments and in a real mobile
manipulator. The learning process is analyzed with other 2D learning
approaches using the LASA dataset and it is tested in complex 3D
scenarios with different obstacles, proving its effectiveness.