HERE: Hierarchical Active Exploration of Radiance Field with Epistemic Uncertainty Minimization

TL;DR

HERE tackles efficient, high-fidelity 3D scene reconstruction with neural radiance fields by explicitly modeling epistemic uncertainty through evidential deep learning and integrating this signal into a hierarchical planning framework. A grid-based UQ module provides fast, locally updatable uncertainty estimates, while a global planner ensures complete coverage and a local planner targets informative viewpoints for detailed reconstruction. The approach achieves state-of-the-art completion on photorealistic simulated scenes and is validated with real-world hardware demonstrations, demonstrating practical viability for robotics and perception tasks. Overall, the method enables real-time, scalable mapping by coupling uncertainty-aware exploration with neural implicit mapping.

Abstract

We present HERE, an active 3D scene reconstruction framework based on neural radiance fields, enabling high-fidelity implicit mapping. Our approach centers around an active learning strategy for camera trajectory generation, driven by accurate identification of unseen regions, which supports efficient data acquisition and precise scene reconstruction. The key to our approach is epistemic uncertainty quantification based on evidential deep learning, which directly captures data insufficiency and exhibits a strong correlation with reconstruction errors. This allows our framework to more reliably identify unexplored or poorly reconstructed regions compared to existing methods, leading to more informed and targeted exploration. Additionally, we design a hierarchical exploration strategy that leverages learned epistemic uncertainty, where local planning extracts target viewpoints from high-uncertainty voxels based on visibility for trajectory generation, and global planning uses uncertainty to guide large-scale coverage for efficient and comprehensive reconstruction. The effectiveness of the proposed method in active 3D reconstruction is demonstrated by achieving higher reconstruction completeness compared to previous approaches on photorealistic simulated scenes across varying scales, while a hardware demonstration further validates its real-world applicability.

Figures (8)

• Figure 1: Our approach performs active scene reconstruction of neural radiance fields by leveraging epistemic uncertainty to guide exploration of an agent. It identifies regions with high epistemic uncertainty and generates an informative path. This, combined with hierarchical planning, ensures both scene coverage and detailed exploration of uncertain areas, enabling high-quality reconstruction across various scales.
• Figure 2: An overview of the framework. Given RGB and depth images, the neural implicit SLAM module learns an implicit map along with evidence and sufficient statistics grids, which are used to quantify epistemic uncertainty. The hierarchical active reconstruction planning method leverages this uncertainty for target view selection in local planning and frontier extraction for global region path generation. The planner then generates a camera trajectory to explore the unseen environment.
• Figure 3: Illustration of our uncertainty quantification module. The spread around $(\mu,\sigma)$ enlarges in poorly reconstructed regions, which we model as epistemic uncertainty.
• Figure 4: Illustration of the process of the scene coverage planning and local trajectory planning.
• Figure 5: (a) Uncertainty and SDF prediction error of the ground truth mesh on scenes from the Gibson dataset (Top: Pablo, Bottom: Swormville) for each algorithm during exploration. Each value is normalized, with red (1) indicating a high value and blue (0) indicating a low value. (b) AUSE plot for the UQ metrics evaluated on selected scenes from the Gibson dataset.