A novel MPC approach to optimize force feedback for human-robot shared control

摘要

One of the challenging problems in human-robot shared control is the algorithms for force rendering due to uncertain human behavior. Common adaptive and optimal control techniques may not be readily applied to a number of popular haptic devices, since the required state space models are not available. In addition, the invoked control algorithms should meet the requirements of a human-in-the-loop control problem and be fast enough for such haptic applications. Thus, this paper proposes a novel combination of model predictive control (MPC) and neural networks to overcome the aforementioned problems. First, the robot is modeled by a multilayer perceptron (MLP) network. The model has been trained with a set of data created with reasonably considered all possible robot variations in a virtual environment (VE). Afterwards, an MPC is developed with the aid of artificial neural networks and genetic algorithm (GA) to find the optimized force required for a typical task model. Such optimal force calculation with the MPC approach has been carried out for a variety of reference points of a task. Then another MLP network is trained to find the optimal required forces for a reasonably fast and smooth performance. The results show the effectiveness of this approach compared to classical impedance controllers in increasing user performance.