RT-GuIDE: Real-Time Gaussian Splatting for Information-Driven Exploration

TL;DR

RT-GuIDE introduces a real-time active mapping and exploration framework based on 3D Gaussian Splatting (3DGS) and a GPU-accelerated planner that operates onboard to enable information-driven exploration. The mapping component builds a dense Gaussian map and estimates uncertainty directly from Gaussian parameters, while the planning component uses a hierarchical scheme (high-level region guidance and low-level trajectory optimization) to maximize information gain along feasible paths. The approach achieves high-quality map reconstructions with photometric and geometric fidelity, provides an 18x speedup in planning over CPU baselines, and demonstrates superior performance in both simulated MP3D scenarios and real-world experiments, including semantic segmentation with rendered views. The work highlights practical benefits for real-time autonomy and provides ablations, benchmarks, and open-source release potential for broader adoption in information-driven robotic exploration.

Abstract

We propose a framework for active mapping and exploration that leverages Gaussian splatting for constructing dense maps. Further, we develop a GPU-accelerated motion planning algorithm that can exploit the Gaussian map for real-time navigation. The Gaussian map constructed onboard the robot is optimized for both photometric and geometric quality while enabling real-time situational awareness for autonomy. We show through viewpoint selection experiments that our method yields comparable Peak Signal-to-Noise Ratio (PSNR) and similar reconstruction error to state-of-the-art approaches, while being orders of magnitude faster to compute. In closed-loop physics-based simulation and real-world experiments, our algorithm achieves better map quality (at least 0.8dB higher PSNR and more than 16% higher geometric reconstruction accuracy) than maps constructed by a state-of-the-art method, enabling semantic segmentation using off-the-shelf open-set models. Experiment videos and more details can be found on our project page: https://tyuezhan.github.io/RT GuIDE/

Figures (5)

• Figure 1: Key elements of our proposed approach. [A] Robot building a Gaussian map onboard in real-time and using it to avoid obstacles in the environment. Synthesized color and depth images from the Gaussian map are presented next to the corresponding observations from the RGBD sensor. [B] Robot navigating to unobserved areas (right) with high information gain while maximizing information along the trajectory. Map regions colored light to dark in increasing estimated information gain.
• Figure 2: The proposed active mapping framework. The proposed framework contains two major components, the planning module and the mapping module. As can be seen in the figure, the Mapping module ([A]) takes in RGB, depth and pose measurements, and updates the map representation at every step and computes the utility of cuboidal regions (Sec. \ref{['sec:high-level-guidance']}). The utility of each region is then passed to the planning module which comprises the topological graph and motion primitive library ([B]). The trajectory planner in turn attempts to plan a path to goal that maximizes information gathering (queried from the mapper). The planned trajectory is executed by the robot to get a new set of observations.
• Figure 3: Time spent to plan a trajectory with a 5m horizon (including 129 collision checking points) versus the number of Gaussians in the map.
• Figure 4: Visualization of onboard constructed map
• Figure 5: Qualitative results