LGSDF: Continual Global Learning of Signed Distance Fields Aided by Local Updating

TL;DR

This work addresses incremental ESDF map construction for robotics by identifying conflicts in training data when using purely implicit representations. It introduces LGSDF, a three-module pipeline that uses front-end active sampling and local grid updating to fuse distances into axis-aligned grids and a back-end self-supervised global learning phase that trains a neural implicit ESDF map. Key contributions include a novel irregularity-based pixel sampling strategy, grid-based local fusion for plasticity and stability, and a loss suite combining SDF, gradient, and Eikonal terms that yield accurate ESDF maps and clean meshes across multiple scenes. The approach delivers lightweight memory usage and real-time performance, with publicly available code, and demonstrates superior accuracy and mesh quality compared with state-of-the-art explicit and implicit methods.

Abstract

Implicit reconstruction of ESDF (Euclidean Signed Distance Field) involves training a neural network to regress the signed distance from any point to the nearest obstacle, which has the advantages of lightweight storage and continuous querying. However, existing algorithms usually rely on conflicting raw observations as training data, resulting in poor map performance. In this paper, we propose LGSDF, an ESDF continual Global learning algorithm aided by Local updating. At the front end, axis-aligned grids are dynamically updated by pre-processed sensor observations, where incremental fusion alleviates estimation error caused by limited viewing directions. At the back end, a randomly initialized implicit ESDF neural network performs continual self-supervised learning guided by these grids to generate smooth and continuous maps. The results on multiple scenes show that LGSDF can construct more accurate ESDF maps and meshes compared with SOTA (State Of The Art) explicit and implicit mapping algorithms. The source code of LGSDF is publicly available at https://github.com/BIT-DYN/LGSDF.

Figures (11)

• Figure 1: The demonstration of zero-level set meshes of ESDF constructed by different strategies. Explicit mapping dynamically updates signed distances within the discrete grids and thus has limited resolution and is not predictive for unobserved regions. Implicit mapping regresses the SDF using a simple MLP, but often causes scene distortion due to conflicting training data. In contrast to these, LGSDF reconstructs a continuous, smooth map with sharp object boundaries.
• Figure 2: The framework of LGSDF consists of three main modules: Active Sampling, Local Updating and Global Learning. With the continuous input of the posed depth images flow, the implicit ESDF neural network gradually recovers the full SDF of the observed scene through self-supervised continual learning.
• Figure 3: Active pixel sampling. The irregularity ${\xi}_{b}$ of each block is inferred from the depth image $\mathbb{D}$ and the normal rendering image $\mathbb{N}$ to allocate more sample pixels in more complex regions.
• Figure 4: There are differences in the approximate distances of the same query point in different frames. If these conflicting data are fed directly into the network, the performance of the implicit map will be affected.
• Figure 5: An example of selecting training grids. Left: The current awareness of the scene. Right: The selected grids for training. The Currently updated grids and the sampled historically updated grids guarantee good plasticity and stability of the network, respectively.