Towards Large-Scale Incremental Dense Mapping using Robot-centric Implicit Neural Representation

TL;DR

This work tackles large-scale dense mapping with implicit neural representations by introducing Robot-centric Implicit Mapping (RIM). RIM uses a hybrid representation that stores learnable features in a multi-resolution voxel grid and decodes SDF values via a shallow MLP, trained locally on a robot-centered map while sharing a decoupled global map to maintain memory efficiency. It introduces bundle supervision and spatial sampling with outlier removal to mitigate catastrophic forgetting and dynamic object interference, enabling unbounded, real-time incremental mapping. Across Replica, MaiCity, and KITTI, RIM achieves superior reconstruction quality and efficiency, maintains constant video memory, and demonstrates applicability to real-world robotic tasks with robust performance under noise and dynamics. The approach offers practical implications for robotics, digital twins, and AR/VR scenarios requiring scalable, high-fidelity dense mapping with limited onboard memory.

Abstract

Large-scale dense mapping is vital in robotics, digital twins, and virtual reality. Recently, implicit neural mapping has shown remarkable reconstruction quality. However, incremental large-scale mapping with implicit neural representations remains problematic due to low efficiency, limited video memory, and the catastrophic forgetting phenomenon. To counter these challenges, we introduce the Robot-centric Implicit Mapping (RIM) technique for large-scale incremental dense mapping. This method employs a hybrid representation, encoding shapes with implicit features via a multi-resolution voxel map and decoding signed distance fields through a shallow MLP. We advocate for a robot-centric local map to boost model training efficiency and curb the catastrophic forgetting issue. A decoupled scalable global map is further developed to archive learned features for reuse and maintain constant video memory consumption. Validation experiments demonstrate our method's exceptional quality, efficiency, and adaptability across diverse scales and scenes over advanced dense mapping methods using range sensors. Our system's code will be accessible at https://github.com/HITSZ-NRSL/RIM.git.

Figures (8)

• Figure 1: Pipeline of the robot-centric implicit mapping. RIM generates signed distance fields in real-time based on provided posed points.
• Figure 2: The schematic diagram of spatial sampling. The high-quality input within the perceptual range carves the surface in detail, while the outside input distinguishes the free space. Moreover, the local map dynamically maintains historical points for bundle supervision.
• Figure 3: The schematic diagram of the implicit map structure and the sub-maps management. The local map slides in the global map. Each grid represents a sub-map, and those without color are non-allocated. A hash table arranges the allocated sub-maps. Red grids indicate active sub-maps stored in video memory, and pink grids indicate frozen sub-maps stored in system memory.
• Figure 4: Reconstructed mesh on the Replica dataset's office-2 and room-1, and colors indicate the direction of the surface normal. Our method can capture richer, more complete, and precise geometric details in small objects like tables, pillows, and quilts.
• Figure 5: Reconstructed mesh on the MaiCity dataset sequence 01, and colors indicate the surface normal direction. Our method outputs smoother and more complete mapping results, such as the trees, car, walker, walls, and the ground.