Research On Door Opening Skill Learning Of Tactile Manipulator Based On Deep Reinforcement Learning

With more and more research on tactile sensors,tactile sensors are being used to install on home service robots to improve their service performance.The robot door opening skill is the basic and important skill in the home service task.Therefore,this paper studies the door opening skill learning of tactile robotic arm.Based on the modal information of tactile and ontological perception in the door opening process of the robotic arm,and combined with the algorithm framework built based on deep reinforcement learning,the door opening skill of the tactile robotic arm is improved.Specific research contents are as follows:Firstly,to solve the problem of low experience pool sample utilization rate in deep reinforcement learning in the door opening skill learning of tactile manipulators,a state reward prediction model and meta ACON activation function based double delay algorithm(SRATD)was proposed.The algorithm first draws inspiration from the deep reinforcement learning idea based on models and proposes a state reward value prediction model.This model predicts future states and reward values through data in the experience pool,making the algorithm forward-looking.At the same time,the meta ACON activation function is added to the proposed prediction model to realize the adaptive activation of the activation function,so as to improve the learning ability of the prediction model.Finally,an update formula for the actor network was defined based on the predicted model to determine the future state and reward values,in order to improve sample utilization.Secondly,a Tactile Ontological Perception Multimodal Representation Learning And Noisy Linear Layer Twin Delayed(TPNTD)algorithm is proposed to address the issue of low utilization of multimodal information in the learning of tactile robotic arm door opening skills using both tactile and ontological modal information.The algorithm first proposes a multimodal representation learning module to process tactile modal information and ontological perceptual modal information,obtaining a compact feature representation that makes more effective use of the sampled multimodal data.The multimodal representation learning module is added to the actor network to form a new actor network to complete the output of tactile robotic arm actions.Simultaneously adding noise linear layers to the actor network and critic network to improve the exploration ability and algorithm robustness of the intelligent agent.Finally,the SRATD algorithm and TPNTD algorithm were experimentally validated using the tactile robotic arm door opening skill learning simulation environment built by Mu Jo Co.Analyzing the results of four experimental indicators proves the effectiveness of the algorithm proposed in this paper.