Mobile Robot Localization Based on Gradient Propagation Particle Filter Network

Abstract

In order to solve the problem that the gradient information can’t propagate backward due to the non-differentiability of resampling process in the end-to-end training of Differentiable Particle Filters (DPFs) network model, a particle filter network with gradient propagation is proposed in this article. It uses a differentiable resampling process in the network and is called Gradient Propagation Particle Filter Network(GPPFN). Firstly, an auxiliary noise vector is added to approximate the particle distribution in the resampling process, so that it can be represented by a certain differentiable function. Then, the method is applied to GPPFN model to make the gradient information generated in the training process can be propagated. It can provide the information of previous prediction and update steps for the next step prediction, so as to improve the accuracy of robot localization. In this article, experiments are carried out in the simulation environment of Deepmind lab. GPPFN is used to train the motion model and measurement model respectively, and the training effect is expressed by likelihood degree. The lower the likelihood degree is, the more consistent the training results are. The experiments show that the likelihood loss of motion model and measurement model are reduced by using GPPFN. Finally, The GPPFN model is compared with other methods. The training results showed that compared with the others, the training results of GPPFN has lower error rate and better robustness.

Keywords

• Robot localization
• particle filter
• deep neural network
• state estimation