| So far,robot dynamics modeling is still a hot issue in the scientific community.The robot control method can directly benefit from the model knowledge,achieve the level of accuracy and consistency required by the task,and reduce the burden of the control system.However,the multi-degree of freedom robot is a complex nonlinear system with strong coupling,and its dynamics have uncertainties such as gravity,inertia,Coriolis force,elastic deformation and assembly clearance.Moreover,due to the complexity of most robot structures and the lack of information provided by manufacturers,it is impossible to directly obtain relevant information and accurate parameters,resulting in a serious impact on modeling accuracy.Therefore,this paper conducts in-depth research on the robot dynamics modeling method based on sparse feature learning.In the absence of relevant prior information related to the physical structure and kinematics,it breaks through the problem of accurate dynamics modeling,and optimizes the robot control methods based on the learning model,providing an effective technical approach for the data-driven"modeling-learning-control"complete scheme.The models are classified based on the analysis of the dynamic equation and Jacobian matrix,and study the modeling problem of the robot static model.In view of the instability of linear regression solution in traditional methods,a model learning method based on kernel technique is proposed,and the results are used as a reference to continue in-depth mining.To avoid the dependence of the modeling method on the physical structure and kinematics parameters,a robot gravity model learning algorithm based on sparse norm optimization is proposed.Based on the basis function dictionary structure of the model and the sparse optimization method,the model is directly learned from the motion sample data with the premise of unknown robot structural parameters.The proposed method can eliminate the model error caused by parameter calibration and assembly,and improve the modeling accuracy.In addition,the alternating direction multiplier algorithm is introduced to optimize the learning process and improve the speed of gravity model calibration and update.Due to the excellent performance of the modeling method based on sparse feature learning in the reconstruction of the robot static model,the feasibility of the method in dynamic modeling is continued to explore.The overcomplete dictionary of the robot dynamic model is analyzed,and it is found that the dynamic model dictionary is more complex in structure and larger in scale,which exceeds the computing ability of the general solver.To solve this problem,a customized solver based on ADMM and reweighted L1 minimization technology is proposed to reduce the time and space complexity of the dynamic model observation equation and solve the difficulties caused by the data"dimension disaster"of the general solver.In addition,the continuous differentiable friction model is used to learn the robot joint friction,solve the problem of friction model design,and form a complete dynamic model of the robot in the actual working state.The robot control based on the sparse feature learning model is studied by taking the scene of the robot-human shared workspace as an example.Two issues are generally considered:one is the safety control of the robot when it works independently and effectively responds to the situation that people or obstacles suddenly appear in the given trajectory.The other is the adaptability to external forces in the contact-oriented operation task and whether it can easily be pulled and dragged.Based on the sparse feature learning model,the optimization design method of feedforward low-gain control and model reference control without kinematics calculation is proposed.The saturation gain increment optimization based on ARE is introduced to improve the controller’s response ability to sudden disturbance changes.The reference controller is optimized based on SMO to prevent the adverse effects of joint coupling.The proposed control method helps to improve the security of robots and the consistency of physical interaction and provides an effective control scheme for human-robot cooperation in the scene.The experiments were conducted on the self-developed 7-DOF articulated robot platform.The joint torque prediction experiments with different loads verify the effectiveness of the model learning method,and the comparative experiments with other dynamic modeling schemes show the advantages of the proposed method.The low-gain controller based on the feedforward design of the learning model realizes collision detection without external sensors,and ensures the safety of the robot body and operators in operation.The model reference controller designed based on the learning model realizes force-free control and interactive force following of the robot.The above experiments verify the effectiveness of the sparse feature learning modeling method and the practicability and reliability of the controller designed based on the learning model in the tasks. |
