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

RA-L 2026

Taekbeom Lee^*, Dabin Kim^*, Youngseok Jang, H. Jin Kim

Seoul National University

^*Equal contribution

HERE performs active 3D scene reconstruction using neural radiance fields, guided by epistemic uncertainty to explore unseen regions and generate informative camera trajectories.

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 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.

Video

Method

Overall Pipeline

HERE integrates a neural implicit SLAM module with an epistemic uncertainty-driven hierarchical planner. Given RGB-D images, the mapping module learns an implicit map together with evidence and sufficient statistics grids, which are used to quantify epistemic uncertainty online. The hierarchical active reconstruction planner leverages this uncertainty for target viewpoint selection in local planning, and for frontier extraction in global region-level path generation. The planner then generates a smooth camera trajectory to guide the robot toward unexplored regions.

Overview of the HERE framework pipeline.

Uncertainty Quantification based on Evidential Deep Learning

We quantify epistemic uncertainty by extending evidential deep learning (EDL) to neural implicit mapping. A Normal Inverse-Gamma (NIG) prior is placed over the SDF prediction at each 3D point, and the posterior is updated via two lightweight learnable grids that predict a confidence score and a second moment. This design ensures spatial locality and real-time performance through trilinear interpolation, while keeping the mapping module unchanged. The entropy of the posterior NIG distribution serves as the epistemic uncertainty measure: it is high in poorly observed or unobserved regions and low where observations are sufficient, exhibiting a strong correlation with actual SDF prediction errors.

Illustration of the epistemic uncertainty quantification module.

Hierarchical Planning

To achieve both comprehensive scene coverage and detailed reconstruction, we propose a hierarchical planning framework. The global planner decomposes the environment into fixed-resolution regions, classifies each as unexplored, exploring, or explored based on frontier and uncertainty information, and determines the optimal visitation order by solving a Traveling Salesman Problem. The local planner then refines the global path into a 6DoF camera trajectory by selecting target viewpoints that maximize joint epistemic uncertainty accumulated along the trajectory, using a greedy submodular optimization strategy. When no informative viewpoints remain, a fallback mechanism uses A* on the region connectivity graph to escape local optima and continue exploration.

Illustration of scene coverage planning and local trajectory planning.

Results

Quantitative Results

Uncertainty Quantification

We evaluate our epistemic UQ against other INR-based active reconstruction methods using the AUSE metric. Our uncertainty maps show strong spatial correlation with SDF prediction errors, outperforming all baselines on every evaluated scene in the Gibson dataset.

Uncertainty and SDF error comparison, and AUSE plot on Gibson dataset.

Reconstruction Completeness over Time

We compare reconstruction completeness over exploration steps against state-of-the-art methods on complex multi-room Gibson scenes. Our method achieves superior completion ratio and lower completion error, demonstrating the effectiveness of epistemic uncertainty-driven hierarchical planning.

Completion ratio and completion metric over exploration steps on Gibson dataset.

Per-Scene Reconstruction Quality

We report completion metrics for individual scenes across both the Gibson and Matterport3D datasets. Our method achieves the best completion ratio and completion error in the majority of scenes, demonstrating consistent superiority across environments of varying scale and complexity.

Qualitative Results

Gibson Dataset

Drag the slider to compare reconstructed meshes between Naruto and our method. Our approach produces more complete and accurate reconstructions, effectively avoiding artifacts and empty holes in challenging regions.

HERE

NARUTO

◀▶

Matterport3D Dataset

We further evaluate on five scenes from the Matterport3D dataset, which contains larger and more complex indoor environments. Our method consistently achieves higher completion ratio and lower completion error compared to baselines, demonstrating scalability to large-scale scenes through hierarchical planning.

Drag the slider to compare reconstructed meshes between Naruto and our method on Matterport3D scenes.