Real-world Visual Navigation in a Simulator: A New Benchmark

Henghui Bao, Kiran Lekkala, Laurent Itti

Keywords: Gaussian Splatting, Simulators, 3D reconstruction, Generative models, Visual Navigation

Abstract:

In this paper, we explore advanced techniques in novel view rendering,
particularly Gaussian Splatting, to create a simulator using a
large-scale outdoor dataset. Our simulator,\textit{Beogym}, is
data-driven and built from data collected using a mobile robot. Our
proposed pipeline processes the dataset to obtain an interconnected
sequence of Gaussian splat files. These are then used by an engine to
load appropriate splat files and render image frames during simulation.
\textit{Beogym} offers first-person view imagery, facilitating realistic
training environments that could be used for enhancing and evaluating
the learning capabilities of autonomous agents for visual navigation. It
incorporates a sophisticated motion model and a sequence graph for
seamless querying and loading of different sectors of the environment.
The result closely resembles real-world navigation through smooth
transitions across splat files.

Submission Number: 15