DiffRadar: Differentiable Radar SLAM doubles map consistency at 70 FPS
Radar SLAM with Gaussian fields cuts trajectory errors by over 50% in feature-poor corridors.
Radar SLAM has long struggled with unstable pose estimation and degraded maps in degenerate or dynamic environments due to reliance on discretized heatmap scan matching. DiffRadar, developed by researchers, replaces this with a differentiable, physics-aware Gaussian field representation. The scene is modeled as anisotropic Gaussian primitives, and radar measurements are rendered directly in range-azimuth and Doppler-azimuth spaces through a differentiable forward model. This allows simultaneous optimization of robot pose and scene structure from raw radar data.
Implementation on commodity FMCW radar hardware and evaluation on the Radarize benchmark and a controlled stress-test suite show dramatic gains. DiffRadar achieves substantial reductions in trajectory error—especially in feature-poor corridor motion—while more than doubling map consistency and maintaining real-time performance at 70 FPS. The system robustly handles corridor degeneracy, motion regime transitions, dynamic clutter, and long-horizon loop closures, demonstrating that modeling radar observations directly in the signal domain yields far more robust and consistent radar-only SLAM for mobile platforms.
- DiffRadar models radar observations as differentiable Gaussian fields rather than discrete heatmaps, enabling joint pose and map optimization.
- Reduces trajectory error significantly on the Radarize benchmark, with the largest gains in feature-poor corridor environments.
- Runs at 70 FPS on commodity FMCW radar hardware while more than doubling map consistency over prior methods.
Why It Matters
Enables reliable robot navigation in low-visibility (poor lighting, adverse weather) without cameras or LiDAR, using only low-cost radar.