NExplore: Exploration with Neural Fields for Autonomous Scene Reconstruction.

Area of Science:

• Robotics
• Computer Vision
• Artificial Intelligence

Background:

• Autonomous robots require efficient methods for mapping unknown environments.
• Implicit Neural Representations (INRs) offer a compact way to represent complex 3D scenes.
• Continual learning is crucial for robots to adapt and update maps over time without forgetting previous information.

Purpose of the Study:

• To develop an online active mapping system for autonomous indoor environment reconstruction using INRs.
• To introduce a continual learning paradigm for optimizing neural fields from sequential observations.
• To enable a mobile agent to explore and build a neural map in real-time.

Main Methods:

• A principled continual learning approach balancing generalization and forgetting by maximizing distribution shifts and minimizing reconstruction error.
• A plug-and-play method for quantifying neural map prediction uncertainty via random weight perturbation.
• Experience replay for incremental map updating and a sparse graph structure integrating geometry, appearance, topology, and uncertainty.

Main Results:

• The NExplore system successfully reconstructs indoor environments in real-time.
• Prediction error is progressively reduced through a self-supervised, data-driven approach.
• The integrated graph structure facilitates efficient planning and decision-making for agent navigation.

Conclusions:

• NExplore is the first online active mapping system utilizing coordinate-based INRs.
• The proposed continual learning paradigm effectively manages map optimization during exploration.
• Experimental validation in diverse environments confirms the system's efficacy for autonomous mapping.