Influence of Sensor Noise and Latency on Navigational Safety of Deep-Reinforcement-Learning-based Planners

Abstract

Recently, mobile robots have become important tools in various industries, like logistics, healthcare, and delivery. Deep Reinforcement Learning (DRL) emerged as an end-to-end approach, which maps raw laser scan observations to robot actions and promises flexible and efficient navigation. Various works have incorporated DRL for navigation in highly dynamic environments. However, these approaches are often trained within simulation with optimal assumptions and perfect modeling of observations, which makes the simulationto-reality gap a relevant issue. In this work, we evaluate the impact of noise and delay modules on the performance of DRL-based navigation approaches. Therefore, we introduce noise and delay modules into a 2D simulation environment - arena-rosnav, which makes it feasible to train and test DRL-based approaches under realistic circumstances. Subsequently, we evaluate the approaches with the different noise modules extensively and demonstrate a strong correlation between the level of noise and the success rate. The results could be valuable to design safe DRL-based navigation approaches.