Undefined-behavior guarantee by switching to model-based controller according to the embedded dynamics in Recurrent Neural Network

For robotic applications, its task performance and operation must be guaranteed. In usual robot control, achieving robustness to various tasks as well as controller stability is difficult. This is similar to the problem of the generalization performance of machine learning. Although deep learning is a promising approach to complex tasks that are difficult to achieve using a conventional model-based control method, guaranteeing the output result of the model is still difficult. In this study, we propose an approach to compensate for the undefined behavior in the learning-based control method by using a model-based controller. Our method switches between two controllers according to the internal representation of a recurrent neural network that established the dynamics of task behaviors. We applied our method to a real robot and performed an error-recovery operation. To evaluate our model, we designed a pick--place task, and induced external disturbances. We present results in simulation and on a real robot.