Modeling and learning control system design for robots using deep learning techniques

Abstract

In recent years, research and development in artificial intelligence dramatically impacted fields like health care, agriculture, and manufacturing. In the robotics field, innovation such as recognition, interaction, and manipulation made a breakthrough in "How a robot can help to improve the quality of human life?". Deep learning for real-time applications became viable as computer power and data became more readily available. In the past classical methods of modeling such as Euler-Lagrange method widley have been used to construct accurate dynamic model of the robots, but due to the uncertainty in the parameters and external disturbances, the accurate and generalized analytical model is too complex to build. On the other hand, deep learning algorithms are general non-linear modelling techniques that can learn from input-output data, making them a good choice for modelling and model-based control system design for robots. In this research, recurrent neural network techniques such as Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) are adopted to overcome the issues of modelling. Then, by effectively combining adaptive sliding mode controller (ASMC) with the deep-learning-based (LSTM-based) inverse dynamic model of the robot a model-based control can be developed. . The deep-learningbased controller is applied for trajectory tracking on various scenarios to verify the effectiveness of this approach. The primary strategy of designing a controller with a deep recurrent neural network is to benefit from its memory, resulting in better generalization, accuracy enhancement, and better estimation of time-varying parameters and disturbances. The simulation and experimental findings show that the suggested controller operates adequately on unknown trajectories and disturbances even without tuning settings.