Control of a piecewise constant curvature continuum manipulator via policy search method

Accurate dynamic models are necessary for fast and precise motion control of continuum manipulators, but can hardly be obtained. In this paper, we propose a method for precise position control of piecewise constant curvature (PCC) continuum manipulators using the guided policy search method. The policy network is trained to control the manipulator to reach target position precisely. Firstly, a linear time-varying Gaussian model fitting to sample trajectories is used to approximate the dynamics of the PCC continuum manipulator. Then, the Gaussian controller that induces the trajectory distribution is optimized by iterative LQR based on the fitted model. Finally, the policy is updated by minimizing the KL-divergence between the policy and trajectory distribution. We have simulated the PCC continuum manipulator with two-segment model and three-segment model, respectively. The results show that the linear time-varying Gaussian model can describe the dynamics of the PCC continuum manipulator and the proposed method can achieve desired performance of position control within a couple of iterations.