Research On Obstacle Avoidance Of Field Robot Based On Deep Reinforcement Learning

With the popularization of artificial intelligence technology,robots are gradually developing in the direction of intelligence,and intelligent robots with different functions bring more and more convenience to people.As a branch of the robot field,field robots play an important role in the investigation,detection,military,search and rescue tasks in complex and harsh scenes.These scenes require the ability of field robots to autonomously avoid obstacles.In traditional obstacle avoidance systems,lidar and multi-sensor fusion are often used for environmental sensing.The hardware cost is high,the power consumption is large,and artificial design features and obstacle avoidance strategies are needed.In this paper,the obstacle avoidance system based on deep reinforcement learning is studied.The convolutional neural network is used as the direction control model of the field robot.The image data acquired by the front camera of the field robot is used as input to control the steering of the robot to realize the obstacle avoidance function.This model is low in cost,fast in the calculation,and can continuously improve model performance through self-learning.In the direction control model of the field robot,to shorten the training time of reinforcement learning,deep learning pre-training+ transfer learning+reinforcement learning is studied.This topic first recorded and adopted a data enhancement method to produce a data set containing more than 600,000 images for the direction control model to supervise the learning process.Secondly,a network structure based on a convolutional neural network is designed.A network consisting of deep separable convolution and the full convolutional layer is adopted.Compared with the traditional convolutional neural network,the number of parameters of the model is greatly reduced.Then,the deep learning technology is used to supervise the model,so that the model has certain obstacle avoidance ability.Finally,migration learning and DQN and DDQN algorithms are used to fine-tune the model to further improve the obstacle avoidance performance of the model.In the project,the field robot is set to drive at a constant speed.To realize this function,this paper proposes a deep reinforcement learning algorithm based on the negative feedback idea and combines with the cyclic neural network to realize the speed control system of the field robot.Based on this algorithm.The speed control model allows field robots to learn to travel at a constant speed from scratch by constantly interacting with the environment.The experimental results show that the data samples recorded in this subject can meet the data requirements in the supervision training process,and the direction control model trained by deep learning technology can realize the obstacle avoidance function only by using the monocular camera.It avoids the need for multi-sensor fusion and modeling of the environment in traditional algorithms and overcomes the shortcomings of high hardware requirements,slow operation speed,and poor flexibility.With the deep reinforcement learning algorithm,the robot can learn and learn independently through the environment,which solves the problem that the deep learning algorithm requires a large number of high-quality data sets and high labor costs,and its obstacle avoidance capability also has a great improvement.Besides,the speed control method proposed in this paper avoids the traditional speed control method to model the system,and the cyclic neural network model with dual input improves the traditional feedforward neural network resistance compared with the traditional neural network model.The shortcomings of poor anti-interference capability and higher control accuracy.