Super4DR: 4D Radar-centric Self-supervised Odometry and Gaussian-based Map Optimization

TL;DR

<3-5 sentence high-level summary> Super4DR tackles robust localization and mapping in adverse environments by centering 4D radar data. It couples a cluster-aware, multi-level self-supervised odometry network with a Gaussian-based map optimizer that uses depth-guided ground completion, geometry-aware densification, and multi-view regularization to reconstruct dense, image-renderable maps from sparse radar points. The approach delivers significant gains in self-supervised odometry, narrows the gap to supervised methods, and achieves superior radar map quality and rendering fidelity compared to prior Gaussian-based methods. The framework demonstrates cross-modal applicability, including radar-thermal reconstruction, and is poised to enhance real-world autonomous systems operating under poor visibility conditions.

Abstract

Conventional SLAM systems using visual or LiDAR data often struggle in poor lighting and severe weather. Although 4D radar is suited for such environments, its sparse and noisy point clouds hinder accurate odometry estimation, while the radar maps suffer from obscure and incomplete structures. Thus, we propose Super4DR, a 4D radar-centric framework for learning-based odometry estimation and gaussian-based map optimization. First, we design a cluster-aware odometry network that incorporates object-level cues from the clustered radar points for inter-frame matching, alongside a hierarchical self-supervision mechanism to overcome outliers through spatio-temporal consistency, knowledge transfer, and feature contrast. Second, we propose using 3D gaussians as an intermediate representation, coupled with a radar-specific growth strategy, selective separation, and multi-view regularization, to recover blurry map areas and those undetected based on image texture. Experiments show that Super4DR achieves a 67% performance gain over prior self-supervised methods, nearly matches supervised odometry, and narrows the map quality disparity with LiDAR while enabling multi-modal image rendering.

Figures (20)

• Figure 1: Comparison of LiDAR and 4D radar data. Unlike LiDAR, 4D radar points usually suffer from greater sparsity, noise, and uneven density, thereby posing challenges for robust odometry estimation and high-integrity mapping. Here, the maps are constructed based on ground-truth poses provided by the NTU4DRadLM ntu4dradlm dataset.
• Figure 2: Overview of Super4DR framework with end-to-end radar odometry and gaussian-based map optimization. First, a cluster-aware network, trained with multi-level self-supervised losses, processes radar points along with cluster grouping to estimate poses. Afterwards, an initial map with incomplete structures is constructed based on the trajectory and painted through image pixels. We further propose a map optimizer adopting gaussians as an intermediate representation, with a radar-specific growth strategy, selective separation and multi-view regularization for complete map reconstruction and detailed image rendering.
• Figure 3: Framework of end-to-end radar odometry with multi-level constraints. The odometry network consists of a point-cluster feature encoder and an ego-motion decoder. For self-supervised training, the loss functions first include cluster-weighted distance and column-wise occupancy comparison, which account for the distribution and noise properties of radar points. Network learning is further selectively guided via soft labels generated by a geometry-based algorithm. We also facilitate discriminative feature extraction using feature contrast, while enforcing motion smoothness through a constant-acceleration assumption.
• Figure 4: Details of cluster feature generator. It takes point features with cluster labels as input to produce per-point cluster features.
• Figure 5: Overview of gaussian-based map optimization. We first initialize a sparse 4D radar map as gaussians and perform attribute optimization by gradients derived from multi-view image rendering with the depth and normal maps. During optimization, we adopt ground completion and geometry-aware densification to grow the number of gaussians to reconstruct missing structure, while employing selective separation to decouple sky floaters and avoid excessive pruning. Finally, the optimized gaussians are restored into a dense radar map.