Instruction Analysis Of Home Service Robot Based On Reinforcement Learning

Enabling service robots understand and execute users’ voice commands is a popular direction in the development of service robots.For this,how to parse the instructions into structured action sequences and map the action sequences to the actions in the robot’s instruction set are two important components.Currently,the commonly used methods for the purpose of extracting structured action sequences in instructions include mainly grammar rule-based methods and deep learning-based methods.However,grammar rulebased approaches do not perform well in handling structured instructions,and as the instructions issued by users to the robot become more diverse and colloquial during the use of the service robot,the predefined grammar rules do not handle these instructions well.Deep learning-based approaches are able to handle a variety of complex instructions with the powerful feature extraction capability of deep neural networks,but the performance of deep neural networks is related to the complexity and parameter size of the network,and deep neural networks require a large amount of labeled data for training,which can also greatly increase the hardware cost of service robots and the training cost of action sequence extraction models in robots.Therefore,a well-performing and lightweight instruction parsing model is the key to solve the current problems in the development of voice instruction-based service robots.First,in order to extract structured action sequences from instructions,we propose a action sequence extraction model based on deep reinforcement learning.Because deep reinforcement learning agent can continuously explore themselves in the working environment to obtain a large amount of more widely distributed training data,the model used in this paper uses a lightweight neural network model and greatly reduces the dependence on the size of the labeled data set in the condition of ensuring the effect of action sequence extraction.Second,in order to map the extracted action sequences into actions that can be performed by the robot,a action sequence mapping model based on bidirectional LSTM is proposed in this paper.In order to enable the model understand the instruction text better,we pretrain the model and adds noise to the input vector of the model using adversarial training to improve the robustness and generalization ability of the model.Besides these,in order to solve the problem that the words in the user’s instructions may not in the robot’s vocab,we choose to use char-level word embedding technology.Finally,in this paper,the PR2 robot is simulated in the ROS-based Gazebo simulation environment,and the robot’s joint movements are planned uniformly using the Moveit!planner to observe the robot’s understanding and execution of the command.