基于Transformer网络的双臂机器人模仿学习方法

A Imitation Learning Method of Dual-arm Robot based on Transformer Network

针对机器人模仿学习示教样本数据量小、学习得到的动作策略泛化能力弱等问题,笔者以双臂机器人为研究对象,提出一种基于Transformer网络的模仿学习方法。该方法首先通过多次迭代增强样本数据,然后基于Transformer网络学习新的动作策略数据。仿真结果表明,借助神经网络的学习能力和泛化能力,该方法实现了机械臂动作的模仿学习,具有一定的泛化性,且与同类算法相比具有较高的精度。

Aiming at the problems of small amount of sample data for robot imitation learning and weak generalization ability of the learned action strategy,the author takes the dual-arm robot as the research object and proposes an imitation learning method based on Transformer network.This method first enhances the sample data through multiple iterations,and then learns new action strategy data based on the Transformer network.The simulation results show that with the help of the learning ability and generalization ability of the neural network,the method realizes the imitation learning of the manipulator action,has a certain generalization,and has a higher accuracy compared with similar algorithms.