HGS-Planner: Hierarchical Planning Framework for Active Scene Reconstruction Using 3D Gaussian Splatting

TL;DR

This work tackles real-time, high-fidelity active scene reconstruction in unknown environments by introducing a hierarchical planning framework that leverages 3D Gaussian Splatting (3DGS). A dual-module approach combines a 3DGS-based representation with adaptive global and local planning, augmented by Fisher Information-based quality gains and a voxel-integrated coverage metric to guide viewpoint selection. The method yields efficient, high-quality reconstructions in both simulated MP3D environments and real-world robot experiments, outperforming NeRF-based methods and prior 3DGS planning approaches. This framework offers a practical path toward real-time, reliable perception for autonomous exploration tasks such as search-and-rescue, with potential extension to swarm robotics in large-scale scenes.

Abstract

In complex missions such as search and rescue,robots must make intelligent decisions in unknown environments, relying on their ability to perceive and understand their surroundings. High-quality and real-time reconstruction enhances situational awareness and is crucial for intelligent robotics. Traditional methods often struggle with poor scene representation or are too slow for real-time use. Inspired by the efficacy of 3D Gaussian Splatting (3DGS), we propose a hierarchical planning framework for fast and high-fidelity active reconstruction. Our method evaluates completion and quality gain to adaptively guide reconstruction, integrating global and local planning for efficiency. Experiments in simulated and real-world environments show our approach outperforms existing real-time methods.

Figures (6)

• Figure 1: In a simulated complex house scene, we implemented our high-fidelity active reconstruction system on a mobile robot equipped with an RGB-D sensor. The colored curves represent the robot's executed trajectories. We showcase the reconstruction results, which include the entire rendered scene, detailed renderings from three different views, and the variation in information gain at a specific view.
• Figure 2: An overview of our efficient autonomous reconstruction system with high-fidelity. Utilizing 3DGS for scene representation, the unobserved areas and the Fisher Information from the GS map are provided in real-time to evaluate the quality and completeness of the online reconstruction. Our proposed active reconstruction planning framework efficiently guides the robot to acquire new scene data, ensuring a comprehensive and high-fidelity 3DGS reconstruction.
• Figure 3: A 3D illustration of pixel-level coverage gain evaluation. Given a set of reconstructed Gaussians and a viewpoint, the coverage gain is rendered by weighting unobserved voxel Gaussians with the transmittance of reconstructed Gaussians along the optical ray.
• Figure 4: The top images are depth maps: the left is the ground truth, and the right is the rendered depth. Below, the left image shows the squared error, and the right illustrates the quality gain.
• Figure 5: Trajectories and the reconstruction results from the top view. Left: Ours, Right: GS-Planner