Learning Kinematics from Direct Self-Observation using Nearest-Neighbor Methods

Commonly, the inverse kinematic function of robotic manipulators is derived analytically from the robot model. However, there are cases in which a model is not a priori available. In this paper, we propose an approach that enables an autonomous robot to estimate the inverse kinematic function on-the-fly directly from self-observation and without a given kinematic model. The robot executes randomly sampled joint configurations and observes the resulting world positions. To approximate the inverse kinematic function, we propose to use instance-based learning techniques such as Nearest Neighbor and Linear Weighted Regression. After learning, the robot can take advantage of the learned model to build roadmaps for motion planning. A further advantage of our approach is that the environment can implicitly be represented by the sample configurations. We analyze properties of this approach and present results obtained from experiments on a real 6-DOF robot and from simulation. We show that our approach allows us to accurately control robots with unknown kinematic models of various complexity and joint types.