Obstacle avoidance navigation method for robot based on deep reinforcement learning

Abstract

Aiming at the navigation problem of mobile robots indoor environment, the traditional navigation algorithm based on D3QN has some problems such as sparse reward and slow training speed of the neural network. This paper proposes a deep reinforcement learning Algorithm (LN-D3QN) based on the D3QN network to realize collision-free autonomous navigation of mobile robots. To improve the efficiency of mobile robot learning and exploration, the Vision Sensor is used to acquire the input data from the environment, and the layer normalization method is used to normalize the input data. An improved reward function is designed, which improves the reward value of the algorithm, optimizes the state space, and alleviates the problem of sparse reward to some extent. The data is stored in a priority experience replay pool, and the network is trained using small batches of data. In addition, we evaluate our method by experiment related to indoor navigation. The experiments show that the robot trained by the improved D3QN algorithm can adapt to the unknown environment more quickly than the basic D3QN algorithm. The network's convergence speed is also improved, and it can complete the obstacle avoidance navigation task more efficiently.