Kinematic control of a redundant manipulator using inverse-forward adaptive scheme

The inverse kinematic (IK) relationship of a manipulator is a one-to-many map which can not be learnt using a feed-forward neural network (FFN). The usual method is to learn the forward kinematic relationship using a FFN and then obtaining the IK solution through inversion of the approximate forward model. The accuracy of the inverse kinematic solution thus obtained is limited by the accuracy of the forward model. However, it is not possible to learn the forward kinematic map accurately even with a large network size and a large training data set. Moreover, one needs to resolve redundancy while solving IK in order to make use of available dexterity effectively. These limitations of a conventional network inversion-based scheme are overcome by using an inverse-forward adaptive scheme where the idea is to learn a local solution around the current joint angle configuration rather than trying to learn the global solution through an accurate forward model. It is shown that the accuracy of the IK solution thus obtained, is almost independent of the accuracy of the forward model. The redundancy is resolved by using a KSOM-SC network architecture as a hint generator for initializing the inverse-forward adaptive scheme. The efficacy of the proposed scheme is demonstrated through simulations and real-time experiment on a 7 DOF PowerCube manipulator. A simple obstacle avoidance task is performed to demonstrate the redundancy resolution process.