Neural Radiance Field based Simulator

Abstract

A NeRF is a neural network trained to replicate input views of a single scene via rendering loss. It directly maps spatial location and viewing direction to color and opacity, effectively acting as a “volume”. This allows for a differentiable rendering of new views. The project aims to utilize NeRFs to construct indoor real-world environments, which can then be navigated using a quadrotor in augmented reality.

Results