A Proposal for an Autonomous Navigation Learning Method via Neural Radiance Fields that Does Not Require Actual Robot

Abstract

This paper investigates using Neural Radiance Fields (NeRF) to enable autonomous navigation simulations without the need for actual robots. NeRF's strength lies in its ability to render photorealistic images, promising a solution to the long-standing challenge of the domain gap between simulation and real-world environments. We present findings that validate the effectiveness of a NeRF-simulated environment for training a reinforcement learning policy. Once trained in the NeRF environment, this policy can navigate an actual robot in the real world.