Research On Manipulator Imitation Learning Based On Meta-learning

At present,with the rapid development of industry,manipulators have been widely used in intelligent factory,aerospace,medical surgery and other fields.Therefore,there is a huge demand for intelligent manipulators that can perform complex tasks and have intelligent decision-making functions.Although large-capacity models such as deep neural networks can learn complex tasks and skills,it often requires a lot of time and data from scratch.Meta-learning and imitation learning enable the manipulator to achieve fast learning by imitating the behavior of expert examples.Therefore,exploring how to combine meta-learning with simulated learning is of great significance and value in theory and application for the new task of fast learning of manipulators.The main research work and achievements of this paper are as follows:(1)Combining with the characteristics of meta-learning,a memory weight integration item suitable for meta-learning algorithm is proposed.By adjusting the plasticity of neurons,the manipulator can learn to learn more effectively in the process of learning multi-task and improve the forgetting problem of multi-task learning.In this paper,the memory weight integration item is explained from the perspective of probability.The performance of the integration item is verified by using a few sample image classification tasks and compared with other regularization items.Experiments show that the integration item of memory weight strengthens the learning ability of meta-learning,and the learning effect is obviously better than other regularization items;(2)In meta-imitation learning,memory weight integration term is added and convolution neural network is combined to improve meta-imitation learning algorithm.A deep meta-learning model is constructed.In this paper,layer normalization strategy,Xavier initialization and Adam algorithm are used to optimize the model.In the OpenAI gym Pusher simulation experiment of a 7-DoF manipulator,the influence of integration coefficient on the performance of the model is explored.The success rate of the manipulator's task is counted to evaluate the performance of the model.The simulation results show that the task success rate of the improved meta-simulation learning algorithm is significantly improved when the integration coefficient is 0.6;(3)In order to learn new tasks from the video demonstration of the manipulator and remove the supervision of the manipulator's behavior,an adaptive objective function is constructed based on the time convolution network to provide appropriate gradient information for the strategy and further improve the structure of the deep meta-learning model.SELU activation function and Dropout layer are used to optimize the model.The improved model evaluates its learning performance in OpenAI gym Pusher experiment and compares it with the meta-learning algorithm based on two-head architecture.Experiments show that the improved deep meta-learning model can learn new tasks efficiently only from the video demonstration of the manipulator.Compared with the two-head meta-learning algorithm,it greatly improves the success rate of tasks.