Force and state estimation and control in robotic hand of Surena IV based on limited measurements

摘要

In this paper, an alternative solution is proposed for robotic hands force control by using optimal estimators and controller. This approach does not rely on high cost sensory setup for force sensors. As we considered least cost sensors for a robotic hand (rotary encoder and current sensor), we estimate both grasp force and full states simultaneously using dual Kalman filter algorithm. The dual Kalman filter used in this paper does not have the observability and rank deficiency problem which exist is in augmented state parameter formulation. For control we consider two approaches; first is control through Deep Deterministic Policy Gradient (DDPG) which is an actor critic based reinforcement learning algorithm, this network capture robotic hand experience during different trials and trains actor and critic networks for maximizing accumulative reward in every episode. The control method does not rely on dynamic modelling and can model uncertainty within the networks. Second approach is classical Linear Quadratic Regulator (LQR), which is an optimal state feedback controller. Both of the controllers make the hand follow different reference forces with 0.1% error in 0.3 second.