Gaussian Splatting as a Unified Representation for Autonomy in Unstructured Environments

In this work, we argue that Gaussian splatting is a suitable unified representation for autonomous robot navigation in large-scale unstructured outdoor environments. Such environments require representations that can capture complex structures while remaining computationally tractable for real-time navigation. We demonstrate that the dense geometric and photometric information provided by a Gaussian splatting representation is useful for navigation in unstructured environments. Additionally, semantic information can be embedded in the Gaussian map to enable large-scale task-driven navigation. From the lessons learned through our experiments, we highlight several challenges and opportunities arising from the use of such a representation for robot autonomy.

We adopt the traditional mapping and planning pipeline for autonomy. The mapping module integrates visual and pose measurements into the Gaussian splatting map. A task-dependent metric is used to identify areas in the map with high utility. The high-level planning module computes a path based on cost-benefit analysis of utility and distance for the robot to follow. The low-level planning module determines safe trajectories through the Gaussians. An overview of such a system is presented in the figure above.

Unlike traditional mapping representations, the Gaussian map does not explicitly encode occluded or free space. Instead of relying on frontiers of unobserved regions, we propose in RT-GuIDE an efficient metric for estimating uncertainty in the Gaussian splatting map to determine areas that should be explored next.

Our work in ATLAS Navigator shows that the Gaussian splatting representation can be used to address the challenge of task-oriented navigation. In particular, we demonstrate that we can use Gaussian splatting to incrementally build a rich semantic map and simultaneously identify regions in the map with high relevance for task-driven navigation.