cs.CVMar 19, 2026
SaveDROID-SLAM in the Wild
Organizations: ETH Zurich · Microsoft
Abstract
We present a robust, real-time RGB SLAM system that handles dynamic environments by leveraging differentiable Uncertainty-aware Bundle Adjustment. Traditional SLAM methods typically assume static scenes, leading to tracking failures in the presence of motion. Recent dynamic SLAM approaches attempt to address this challenge using predefined dynamic priors or uncertainty-aware mapping, but they remain limited when confronted with unknown dynamic objects or highly cluttered scenes where geometric mapping becomes unreliable. In contrast, our method estimates per-pixel uncertainty by exploiting multi-view visual feature inconsistency, enabling robust tracking and reconstruction even in real-world environments. The proposed system achieves state-of-the-art camera poses and scene geometry in cluttered dynamic scenarios while running in real time at around 10 FPS. Code and datasets are available at https://github.com/MoyangLi00/DROID-W.git.
Figures & tables
Figure 2: System Overview. The proposed DROID-W takes a sequence of RGB images as inputs and simultaneously estimates camera poses while recovering scene geometry. It alternatingly performs pose-depth refinement and uncertainty optimization in an iterative manner. The proposed uncertainty-aware dense bundle adjustment weights reprojection residuals with per-pixel uncertainty to mitigate the influence of dynamic distractors. In addition, we use predicted monocular depth as regularization of bundle adjustment, to improve its robustness under highly dynamic environments. For the uncertainty optimization module, we first extract DINOv2 [ 37 ] features from the input images and then iteratively update the dynamic uncertainty map by leveraging multi-view feature consistency. Specifically, feature consistency is measured by the cosine similarity between features of image and its corresponding features in image , where the rigid-motion correspondences are derived using the current pose and depth estimates.
Figure 3: Uncertainty Estimation. WildGS-SLAM [ 66 ] and our approach estimate dynamic uncertainty, whereas MonST3R [ 62 ] predicts a binary motion mask. Our approach produces more accurate and spatially consistent uncertainty estimations across all challenging sequences.
Table 1: Tracking Performance on the Bonn RGB-D Dynamic Dataset [ 38 ] (ATE RMSE [cm]). Best results are highlighted as first , second , and third . “-” indicates sequences without reported results in the original papers or unavailable code. “F” denotes tracking failure. For MonST3R [ 62 ] , we use the same keyframes as our method and perform evaluation in a window-wise manner with a window size of 20 and an overlap ratio of 0.5 to reduce memory consumption.
Table 2: Tracking Performance on TUM RGB-D Dataset [ 46 ] (ATE RMSE [cm]). Best results are highlighted as first , second , and third . “-” indicates sequences without reported results in the original papers or unavailable code. Our approach consistently leads to the best or second-best results on the sequences, on average, outperforming all baselines.
Table 3: Tracking Performance on DyCheck Dataset [ 11 ] (ATE RMSE ). Best results are highlighted as first , second , and third . “-” indicates sequences without reported results in the original papers or unavailable code. “F” means tracking failure. Our approach demonstrates the effectiveness and robustness in highly-textured, diverse environments, where prior methods relying on object segmentation or Gaussian mapping for uncertainty optimization often fail.
Table 4: Tracking Performance on DROID-W Dataset (ATE RMSE [m]). Best results are highlighted as first , second .
Table 5: Runtime Comparisons (average FPS ). All evaluations are conducted on an RTX 3090 GPU with a 16-core CPU.
Figure 4: 3D Reconstruction Comparisons on YouTube Sequences. We compare 3D reconstruction quality of DROID-SLAM [ 48 ] , WildGS-SLAM [ 66 ] , and our method. Point clouds from DROID-SLAM and ours are visualized directly, while Gaussian renderings from WildGS-SLAM are displayed using the 3DGS viewer. WildGS-SLAM fails on most sequences. DROID-SLAM shows obvious scale drift ( St. Moritz 1 ), inaccurate geometry ( St. Moritz 3 ), and noisy distractors ( Tokyo Walking 2 & 3 ) under challenging dynamic environments. Our approach produces accurate and consistent reconstructions across highly dynamic and visually challenging real-world sequences.
| Method | ATE RMSE [cm] |
| a. w/o Uncertainty-aware BA | 5.13 |
| b. w/o monocular depth | 3.30 |
| c. w/o uncertainty decouple | 2.57 |
| d. w/o affine mapping | 2.47 |
| e. w/o weight decay | 2.34 |
| Full | 2.30 |
Table 6: Ablation Studies on Bonn RGB-D Dataset [ 38 ] . Details about each configuration are described in Sec. 4.2 .
| Sequence | Number of Frames | Length of Trajectory [m] |
| Downtown 1 | 1427 | 0 90.83 |
| Downtown 2 | 2200 | 122.25 |
| Downtown 3 | 1438 | 0 62.33 |
| Downtown 4 | 1794 | 0 85.19 |
| Downtown 5 | 2157 | 129.93 |
| Downtown 6 | 1900 | 104.99 |
Table 7: Overview of Our DROID-W Dataset.
Table 8: Tracking performance of FAST-LIVO2 [ 65 ] on the DROID-W dataset (ATE RMSE [m]). FAST-LIVO2 is an efficient and accurate LiDAR-inertial-visual fusion system capable of delivering centimeter-level localization accuracy. Its performance on Downtown 3-7 demonstrates sufficient accuracy to serve as ground truth for Downtown 1-2 .
| Sequence | Time | FPS | Resolution |
| Elephant Herd | 00:08 | 24 | 1280 720 |
| Giraffe | 00:09 | 24 | 1280 720 |
| Taylor | 01:13 | 30 | 1280 720 |
| Tomyum 1 | 01:40 | 30 | 1280 720 |
| Tomyum 2 | 01:40 | 30 | 1280 720 |
| St. Moritz | 30:00 | 50 | 1280 720 |
Table 9: Overview of Downloaded YouTube Videos.
Figure 5: Uncertainty Estimation.
Figure 6: Uncertainty Visualization for Consecutive Frames of YouTube Tomyum 1 . We observe that our method robustly handles scenarios in which an object transitions from a static state to dynamic motion, such as a door being pushed open by a person. Prior to the onset of motion, our approach leverages stable visual correspondences on the door to help tracking, since our uncertainty optimization is based on frame-to-frame feature alignment.
Figure 7: Uncertainty Visualization for Consecutive Frames of YouTube Tomyum 2 . Our approach assigns high uncertainty to regions exhibiting strong view-dependent effects (e.g., the mirror).
Figure 8: 3D Reconstruction Comparisons on YouTube Sequences. We compare reconstructed static point clouds between DROID-SLAM [ 48 ] and our method. Our approach produces more accurate and consistent reconstructions across highly dynamic and visually challenging real-world sequences. For the Taylor 22 scene, DROID-SLAM reconstructs two separate corridor structures due to inaccurate pose tracking, as highlighted in the boxed region, whereas our method produces a consistent geometric reconstruction. Moreover, the white lane marking on the asphalt road of Tokyo Walking 1 is faint and fragmented in the point cloud reconstructed by DROID-SLAM, while in our reconstruction it remains cleanly visible, continuous, and highly accurate.
Figure 9: Point Clouds Visualization from 4 Views. Top view : reconstructed static scene from 4 input views shown in the figure. Bottom view : reconstructed dynamic point clouds. The comparisons between the dynamic and static point clouds further demonstrate the effectiveness of our uncertainty estimation. In both visualizations, the static point clouds remain highly consistent, indicating that our method reliably preserves static geometry while filtering dynamic regions.
Figure 10: Qualitative Results of Our Static and Dynamic Reconstruction. We visualize globally aligned static reconstructions alongside dynamic point clouds across all keyframes. Notably, we apply the estimated per-frame dynamic uncertainty to filter out dynamic points. The left and right columns show the static and dynamic reconstructions, respectively. These comparisons highlight the accuracy of our uncertainty estimation, as our method effectively suppresses dynamic regions while preserving geometric fidelity in static areas.
| Method | ATE RMSE [cm] |
| w/o prior term | 5.18 |
| Full | 2.30 |
Table 10: Ablation Studies on Bonn RGB-D Dataset [ 38 ] .