Model predictive control based deep neural network for dynamic manipulation

Abstract

This paper proposes a trajectory planning method with neural networks that learn model predictive control for dynamic manipulation. The novelty of this method is that target positions and model parameters are added to the input of the neural network. The proposed method solves the dynamic trajectory planning issues with low calculation cost. This paper shows the effectiveness of the proposed method through demonstrations of a robot turning a pancake over under various parameters. In addition, we verify a performance change with respect to the number of training data.