Model predictive optimization for imitation learning from demonstrations

Abstract

“Motion generation by imitating” enables a robot to generate its trajectory in a new environment. Research works on dynamic movement primitives (DMP) has reported promising results, with good imitation effect and convergence to the target. However, DMP still has issues such as learning from multiple demonstrations for different initial conditions and achieving obstacle avoidance considering the distribution and motion of obstacles. One of the effective solutions is combining DMP and model predictive control (MPC). The imitation process was transformed into a receding horizon planning procedure, letting the robot to learn more from nearer demonstrations. It is solved as an optimization problem with obstacles modeled as constraints. However, its drawback includes the heavy computation burden, which can be even aggravated in a multi-obstacle scenario where complicated constraints occur. Thus, in this paper, we propose an enhanced MPDMP+ method that combines the advantages of MPC with potential function for both multi-demonstration imitation and multi-obstacle avoidance effect. A proximal augmented Lagrangian method is proposed to solve the optimization problem. This proposed method has a faster convergence rate and small errors. We conducted the simulation and robot experiments for imitation learning for obstacle avoidance scenarios. Our results illustrate the superior performance of the proposed method.