Research On Adaptive Improvement Strategy Of GMM/GMR Towards To Learning From Demonstration(LfD)

Learning from Demonstration(LfD)is a technology that transfers human skills to robots through human-robot interaction and demonstrations.It involves important researches in many robot-related fields such as human-robot interaction,machine learning,machine vision and motion control etc.It provides a very flexible approach to program compared to purely manual programming.When robot collaborates with human,the complexity and variability of the environment requires that the robot not only learn to reproduce the skills demonstrated by human,but also have the ability to perform skills in the new environment.Therefore,from the perspective of complexity of motion and environmental uncertainty,it is important to study LfD.Firstly,oriented to the complexity of motion,the proposed GMM/GMR improvement strategy not only enables aperiodic motion modeling,but also models periodic and quasi-periodic motion.Since the periodic motion has strong spatial-temporal correlativity,an auxiliary algorithm is proposed to solve the spatial-temporal decompling problem.Quasi-periodic motion is a motion that contains periodic characteristics and whose envelope is affected by aperiodic components.Therefore,this paper proposes an algorithm framework to align the moiton modeling,learning and generalization for quasi-periodic motion.The algorithm flow is roughly as follows:equivalent model;decompose quasi-periodic motion into aperiodic and periodic components which are easy to model;model each component separately,and synthesize each component into a quasi-periodic motion.Secondly,oriented to the uncertainty of motion,this paper proposes a motion adaptive adjustment method based on LfD.When robots play collaborative roles with humans in a dynamic environment,they must be able to learn and execute new behaviors to achieve desired tasks.In this pap er,new obstacles(not appearing in the teaching process)are regarded as the environmental uncertainties,so the algorithm aims to avoid the obstacles to complete the original tasks.The algorithm flow is roughly as follows:learn motion distributions;data resampling;data partitioniong and reorganization;motion replanning,and select the most appropriate trajectory according to the optimal strategy.In order to verify the effectiveness of the algorithm proposed in this paper,not only performing simulation experiments in the MATLAB environment,but also performing the physical experiments based on the ABB IRB1200 robot.Both experimental results prove the accuracy and effectiveness of the proposed algorithm.