机器人示教学习与技能泛化实验设计

Experimental design of learning from demonstration and skill generalization for robots

该文提出了一种基于视觉信息和动态运动原语(dynamic movement primitive,DMP)模型的机器人示教学习方法。首先,利用双目摄像头拍摄数据,并基于掩膜原理、质心计算和坐标变换,实现基于视觉信息的场景目标位置识别;其次,基于DMP模型的技能学习算法,提出了初始位置和目标位置相同情况下的分段学习方法;最后,将上述目标识别算法和技能学习方法相结合并应用于机器人作业任务中,通过仿真实验验证了所提出示教学习方法的有效性,实现了机器人在位置空间上的技能泛化。该方法对于机器人示教学习以及人机技能传递的相关实践教学和应用研究具有重要意义。

This paper proposes a robot teaching and learning method based on visual information and dynamic motion primitive(DMP) model. Firstly, the binocular camera is used to capture the data, and based on the mask principle, centroid calculation and coordinate transformation, the scene target position recognition based on visual information is realized. Secondly, based on the skill learning algorithm of DMP model, a segmented learning method with the same initial position and target position is proposed. Finally, the above-mentioned target recognition algorithm and skill learning method are combined to apply to robot task. The effectiveness of the teaching and learning method is verified by simulation experiment, and the generalization of robot skills in position space is realized. This method is of great significance for robot teaching and learning as well as practical teaching and application research on human-computer skills transfer.