A primal-dual neural network for kinematic control of redundant manipulators subject to joint velocity constraints

The paper presents a primal-dual neural network for inverse kinematics computation of redundant manipulators subject to joint velocity limit constraints. While the desired velocities of the end-effector are fed into the neural network, it generates the joint velocity vector which is drift-free and never exceeds the hardware limits. The consideration of actuator limits prevents a manipulator from nonlinear saturation, and hence keeping a good tracking accuracy. The primal-dual network is proven to be asymptotically convergent to exact solutions. Simulation results show that the presented approach is capable of effectively generating the optimal redundancy resolution