Research On Data-driven Industrial Robot Dynamic Identification,Control And Planning

Modern robots are developed to have the capabilities of sensing and adapting to multiple operating environments,which is of great significance in reducing human effort,improving production efficiency,saving energy consumption,and ensuring personal safety.However,the following problems remains unsolved in the development of industrial robots:inaccurate identified dynamic models resulting in unstable joint control,the inefficient energy consumption,and collision problems in human-robot safety are often not considered.This dissertation starts from the basic algorithm framework of industrial robots and considers the accuracy,robustness,and low level of intelligent control strategy of the current industrial robot algorithms.It proposes a data-driven identification and control strategy algorithm framework for industrial robots.Its main contributions are as follows:A sparse Bayesian learning-based online dynamics recognition algorithm for industrial robots is proposed for the accuracy problems caused by the unmodeled terms and system interference in industrial robot dynamics recognition.By combining a priori physical knowledge of robot dynamics and potential candidate mechanism models to form a dictionary function,the nonlinear time-varying dynamics of the robot can be morphed into a linear parametric form.The accuracy of the online identification method is verified by predicting the moments of a real industrial robot through experiments.A deep adaptive control algorithm for industrial robots is proposed to address the unmodeled terms of industrial robot dynamics that are difficult to be described with explicit functions.A deep neural network is used to fit on the residuals of the above identification method to obtain a complete dynamic model as a compensation of the mechanistic model.Subsequently,the stability of the proposed closed-loop system is proved according to Lyapunov’s second theorem,and the adaptive update law of the algorithm is designed accordingly.The effectiveness of the proposed algorithm is also verified by zero-force control and trajectory tracking experiments on an industrial robot platform.To address the problem of energy wastage of industrial robots due to unreasonable trajectory settings,a reinforcement learning-based trajectory planning algorithm for industrial robots is proposed.The designed reward function contains distance,velocity,obstacle avoidance and energy saving terms and a deep reinforcement learning algorithm is used for training.A domain randomization method is used during training in the simulation environment to enhance the efficiency of the training strategy and the subsequent algorithm transferability.Finally,smooth output trajectories are generated by cubic polynomials on an industrial robot platform,and the aforementioned deep adaptive control method is used for trajectory tracking experiments to verify the feasibility of this trajectory planning algorithm and to show a practical reduction in the total energy consumption under a specific task.To address the collision problem in human-robot interaction scenarios,a collision detection technique for industrial robots based on switching momentum dynamics identification is proposed.The collision detection is transformed into a switching system identification problem,and a momentum dynamic is used to avoid the detection inaccuracy caused by the small acceleration signal-to-noise ratio and to quantitatively identify the contact force after the collision.Subsequently,the specific joint location where the collision occurred is accurately located through a classification algorithm retrospectively.The algorithm was first tested in a simulation environment,and then migrated to a home-built industrial robot platform for collision testing.The test results show the practical operability of the proposed algorithm.