Deep reinforcement learning based socially aware mobile robot navigation framework

In this study, we propose a socially aware navigation framework, which enables a mobile robot to avoid humans and social interactions in dynamic social environments, using deep reinforcement learning algorithm. The proposed framework is composed of two main stages. In the first stage, the socio-spatio-temporal characteristics of the humans including human states and social interactions are extracted and projected onto the 2D laser plane. In the second stage, these social dynamic features are then feed into a deep neural network, which is trained using the asynchronous advantage actor-critic (A3C) technique, safety rules and social constraints. The trained deep neural network is then used to generate the motion control command for the robot. To evaluate the proposed framework, we integrate it into a conventional robot navigation system, and verify it in a simulation environment. The simulation results illustrate that, the proposed socially aware navigation framework is able to drive the mobile robot to avoid humans and social interactions, and to generate socially acceptable behavior for the robot.