机器人逆动力学神经网络建模

Study on robot inverse dynamics modeling with neural network

针对神经网络逆动力学模型用于前馈控制的网络泛化能力问题,通过对实际机器人的仿真实验,分析了前馈神经网络学习的机理,对神经网络训练过程中的数据采样环节进行了改进,提出了在状态空间域中进行数据筛选和处理的神经网络学习方法。通过二自由度机器人运动仿真实验表明,该方法提高了模型泛化能力,有利于实时动力学前馈控制方法的实用化;与传统PID控制相比,该神经网络模型进行动力学前馈控制能大幅度减少动态误差,改善了系统稳定性。

Considering the generalization ability of the IDM （ inverse dynamics model） by neural network which is often used in dynamics feedforward control, with plenty of simulation experiments on a robot, the characteristics of neural network learning are discussed. An improving method for neural network training was presented. The improvement aimd at sample data obtaining and disposing, and the method was based on data filtering and processing on state space, The effectiveness of the method was confirmed by simulation experiments on a 2 DOF （ degree of freedom） robot. The experiments show that the generalization ability of the IDM is improved with this method. It is of benefit to utility of real-time dynamics feedforward control. Comparing with PID control, the dynamics feedforward control using such IDM can reduce the dynamic error greatly and improve the stability of the system.