DiffRadar: Differentiable Physics-Aware Radar SLAM with Gaussian Fields

2026-07-13

AI Summary

This work addresses the limitations of traditional radar SLAM, which relies on discrete heatmaps for scan matching, thereby disrupting geometric continuity and failing to accurately model radar physics—leading to unstable pose estimates and degraded mapping quality in feature-sparse or dynamic environments. To overcome these issues, the paper proposes DiffRadar, the first approach to represent radar observations as a differentiable, physics-aware anisotropic Gaussian field. It introduces a differentiable radar forward model that renders measurements in both range–azimuth and Doppler–azimuth spaces, enabling joint optimization of pose and scene structure. By moving beyond conventional discrete matching paradigms, DiffRadar achieves significantly improved robustness and consistency on the Radarize benchmark and stress tests: trajectory errors are substantially reduced, map consistency in degenerate scenarios such as corridors improves by over twofold, and the method supports real-time deployment on FMCW radar at 70 FPS.

Abstract

Radar sensing is increasingly used in mobile systems because it operates reliably under poor lighting, adverse weather, and privacy-sensitive settings where cameras and LiDAR often fail. However, most existing radar SLAM systems estimate motion through scan matching on discretized radar heatmaps, which breaks geometric continuity and fails to capture key radar sensing properties, often leading to unstable pose estimation and degraded mapping in regenerate or dynamically changing environments. We present DiffRadar, a real-time radar SLAM system that models radar observations as a differentiable, physics-aware Gaussian field rather than discrete scans. DiffRadar represents the scene as anisotropic Gaussian primitives and renders radar measurements in range-azimuth and Doppler-azimuth spaces through a differentiable radar forward model, enabling joint optimization of robot pose and scene structure directly from radar measurements. We implement DiffRadar on commodity FMCW radar hardware and evaluate it on both the public Radarize benchmark and a controlled stress-test suite that targets common radar SLAM failure modes, including corridor degeneracy, motion regime transitions, dynamic clutter, and long-horizon loop closures. DiffRadar achieves substantial reductions in trajectory error on the benchmark, with especially large gains under feature-poor corridor motion, while more than doubling map consistency and maintaining real-time performance at 70 FPS. These results show that modeling radar observations directly in the signal domain enables substantially more robust and consistent radar-only SLAM for mobile platforms.

Problem

• Radar SLAM
• discrete radar scans
• geometric continuity
• pose estimation instability
• dynamic environments

Innovation

• Differentiable Radar SLAM
• Gaussian Field Representation
• Physics-Aware Modeling
• FMCW Radar
• Real-time Mapping