Autonomous Learning of Forward Kinematics for Unknown Non-Parallel Manipulators

As any mechanism ages it will undergo structural change. In an autonomous system that operates outside of the lab it is vital that these changes be detected and compensated for. If uncompensated changes are drastic enough the basic motion model will need to be relearned. Most adaptive algorithms require complete relearning when part of the system changes and/or assume only small variations and fail to compensate for more drastic changes. More robust algorithms are needed that do not rely on a priori knowledge to converge to a solution. Manipulator control algorithms use forward kinematics equations to predict the result of an action. The algorithm proposed in this paper is capable of learning the kinematics of a manipulator with very few assumptions. It has the following properties: 1) It can learn the kinematics of an arbitrary non-parallel manipulator with no prior information on the structure or types of joints, 2) it uses multiple tracked features to improve performance, 3) it can identify irrelevant controls and visible features, 4) it operates in real time, 5) fast relearning through reuse of original model, and 6) is capable of learning systems with numerous degrees of freedom (DOF). To the authors knowledge, this is the first such algorithm with all of those features. Simulation results are provided over a wide range of configurations