Mobile Robot Navigation in Dynamic Environments by Offline Reinforcement Learning using Flow-Based Generative Model

Abstract

We propose a novel navigation method applying an offline reinforcement learning method based on Implicit Policy Constraint to mobile robot navigation in environments with pedestrians. The proposed method utilizes a flow-based generative model for the behavior policy, and the latent policy is trained using a method based on Advantage-Weighted Regression. The proposed method is evaluated in a simulation environment.