DROID-W: Geometric Contradiction as a Universal Prior for Dynamic SLAM

Introduction

Simultaneous Localization and Mapping (SLAM) in unconstrained environments remains one of computer vision's persistent challenges. Dynamic objects, from pedestrians and vehicles to swaying vegetation, violate the static world assumption underlying classical geometry, inducing catastrophic drift in camera tracking and corrupting scene reconstruction. Traditional remedies have relied on semantic segmentation networks to mask out predefined object categories, or on geometric heuristics that assume the majority of the scene remains rigid. These approaches falter in complex, cluttered environments where motion is unpredictable and the distinction between foreground and background dissolves.

In DROID-SLAM in the Wild, Li et al. propose a fundamental reframing of this problem. Rather than asking what is moving through semantic classification, their system, DROID-W, asks where the geometry contradicts itself across multiple views. By embedding uncertainty estimation directly into a differentiable bundle adjustment (BA) layer, the method achieves robust tracking and mapping without recourse to object detectors, semantic priors, or assumptions about scene composition. The work achieves state of the art accuracy on dynamic sequences while operating at approximately 10 FPS, and introduces a challenging benchmark of outdoor and YouTube footage that exposes the fragility of existing methods.

Uncertainty Aware Differentiable Bundle Adjustment

The technical core of DROID-W lies in its integration of per pixel uncertainty into the differentiable BA layer inherited from DROID-SLAM. Prior dynamic SLAM systems, such as WildGS-SLAM and UP-SLAM, estimate motion uncertainty through shallow MLPs operating on DINOv2 features. However, these approaches suffer from a critical circular dependency: they require a perfectly static neural implicit or Gaussian Splatting map to supervise the uncertainty predictor. In cluttered real world environments, where dynamic regions may constitute the majority of the frame, constructing such a static map becomes impossible, causing these systems to fail precisely where robustness is most needed.

DROID-W breaks this cycle by deriving uncertainty directly from multi view visual feature inconsistency. During bundle adjustment, the system computes residuals between corresponding features across overlapping frames. Pixels exhibiting high residual error