在人机协作领域,模仿学习是广泛使用的机器人编程方法.在机器人对运动轨迹进行技能学习过程中,针对基于动态运动基元方法(dynamic movement primitives,DMP)只能从单条示教轨迹建模动作的局限性,提出了一种新的将高斯混合模型(gaussian mixture model,GMM)、高斯混合回归(gaussian mixture regression,GMR)和改进的DMP三者相结合的运动轨迹模仿学习方法.该方法先将高斯噪声引入示教轨迹,消除由于示教数据中不可避免存在噪声等不平滑特征带来的不利影响;然后,为了减少单条示教轨迹的不确定性带来的动作建模误差,采用GMM从多条示教轨迹中对运动特征进行编码,使用GMR进行概率轨迹回归;其次,利用改进的DMP算法将轨迹进行泛化,提高机器人技能学习的适应性;最后,通过手写字母轨迹和协作机器人拖动轨迹技能学习试验验证了所提算法的有效性.
Research on Imitation Learning Method of Cooperative Robot Motion Trajectory
In the field of human-machine collaboration,imitation learning is a widely used robotic program-ming method.Aiming at the limitation that the dynamic movement primitives(DMP)can only model the action from a single demonstration trajectory,a new trajectory imitation learning method combining gaussi-an mixture model(GMM),gaussian mixture regression(GMR)and improved DMP is proposed.This method firstly introduces Gaussian noise into the demonstration trajectory to eliminate the adverse effects caused by the inevitable noise and other non-smooth features in the teaching data.Then,in order to reduce the action modeling error caused by the uncertainty of a single demonstration trajectory,GMM is used to encode the motion features from multiple teaching trajectories,and GMR is used for probabilistic trajectory regression.Secondly,the improved DMP algorithm is used to generalize the trajectory in space and im-proved the adaptability of robot skill learning.Finally,the effectiveness of the proposed algorithm is verified by handwritten letter imitation learning simulation and cooperative robot dragging experiment.
万方数据
维普
