NExplore: Exploration with Neural Fields for Autonomous Scene Reconstruction

We present NExplore, an exploration system designed for autonomous reconstruction of indoor environments using implicit neural representations (INRs). The core of the paper is a principled continual learning paradigm that addresses the optimization process of neural fields from sequential observations as an iterative process to achieve a balance between generalization and forgetting. The equilibrium is reached by maximizing the distribution shifts while minimizing the reconstruction error, enabling a mobile agent to actively explore the environment and construct a neural map in real-time, with prediction error gradually reduced through a self-supervised, data-driven fashion. To achieve this, we 1) introduce a plug-and-play method for quantifying the prediction uncertainty of the neural map through random weight perturbation; 2) adopt an experience replay method for incremental map updating; 3) maintain a sparse graph structure that integrates scene geometry, appearance, topology, and uncertainty synergistically for efficient planning and decision-making. With continuous geometric information inherited in the neural map, the agent is guided to navigate traversable paths while progressively enhancing its understanding of the environment. We present for the first time an online active mapping system with coordinate-based implicit neural representations. Experiments conducted in visually realistic Gibson and Matterport3D environments along with real-world scenarios validate the efficacy of the proposed method.