Application Barrier-Free Deep Reinforcement Learning-Based Navigation in Mapless Environment

This paper presents a case study of a novel model-free training framework for indoor mapless navigation without any prior expert demonstrations. A model-free Deep Reinforcement Learning (DRL) approach is implemented, which is capable of good sample efficiency and convergence rate. The navigation policy learns directly end-to-end from raw lidar information in a virtual environment and can be deployed directly on a real mobile robot platform. Without any map or self-location information, the robot is only provided with raw lidar information and the relative position to the target calculated by the inertial measurement unit and encoder which can achieve a large number of success of collision-free navigation tasks during the real-world test. We present a thorough evaluation of different input information, different reward functions, different DRL algorithm structures and different training environments both in simulation and real world. The results show that with appropriate reward function, DRL algorithm structures and environments, the trained agent is able to reach the target without collision. In addition, the policy deployed from virtual to the real robot performs better after processing the input information..