DL-SLAM: Deep Learning Dynamic SLAM

• DL-SLAM is a deep learning-based visual SLAM system that employs Gaussian splatting to model dynamic scenes and retain transiently static objects for pose estimation.
• It utilizes a dual-level probabilistic framework that fuses pixel-level motion estimation with object-level Bayesian updates to enhance tracking and semantic mapping.
• Evaluations on benchmarks like TUM RGB-D and BONN show DL-SLAM achieves improved metrics, reducing RMSE and artifacts compared to traditional dynamic SLAM methods.

DL-SLAM has two distinct uses in recent arXiv literature. In the broad systems sense, “DL-SLAM” denotes deep learning-based visual SLAM systems whose front-end and/or optimization modules are implemented with neural networks rather than only handcrafted features (Pudasaini et al., 2024). In the specific algorithmic sense that now dominates the term, DL-SLAM refers to a monocular Gaussian Splatting SLAM system for dynamic environments whose central objective is to retain useful constraints from transiently static objects for pose estimation while preventing those same objects from contaminating the final map (Xu et al., 2 Jul 2026). The named system is organized around a dual-level probabilistic framework that couples pixel-level dynamic estimation, object-level dynamic probability in 3D Gaussian space, categorical pruning of dynamic Gaussians, and feedback refinement from the purified static map.

1. Terminology and scope

The acronym is not uniform across the SLAM literature. In SPAQ-DL-SLAM, “DL-SLAM” is an umbrella term for deep learning-based visual SLAM systems, with DROID-SLAM used as the concrete reference model (Pudasaini et al., 2024). In contrast, “DL-SLAM” in “DL-SLAM: Enabling High-Fidelity Gaussian Splatting SLAM in Dynamic Environments based on Dual-Level Probability” is the proper name of a specific monocular Gaussian Splatting SLAM method (Xu et al., 2 Jul 2026).

This distinction matters because several similarly named systems denote different methodological families. “DG-SLAM” refers to Differential Geometric SLAM, a geometry-based observer/filter on $SE(3)$ with asymptotic stability claims under ideal assumptions, and is explicitly not a deep learning method (Zlotnik et al., 2015). “DCL-SLAM” denotes a distributed collaborative LiDAR SLAM framework for robotic swarms based on LiDAR-Iris descriptors, distributed loop closure, and distributed pose graph optimization (Zhong et al., 2022). “DF-SLAM” denotes a high-fidelity neural implicit dense visual SLAM system that employs dictionary factors for scene representation (Wei et al., 2024). A common misconception is therefore to read “DL-SLAM” as a generic label for any learning-oriented or similarly abbreviated SLAM pipeline; the literature uses the acronym in both generic and system-specific ways.

Term	Meaning in source	Modality / emphasis
DL-SLAM	Deep learning-based visual SLAM systems; also a named monocular Gaussian Splatting SLAM system	Visual SLAM; either generic or method-specific
DG-SLAM	Differential Geometric SLAM	Geometry-based 3D SLAM on $SE(3)$
DCL-SLAM	Distributed Collaborative LiDAR SLAM	Multi-robot LiDAR SLAM
DF-SLAM	Dictionary Factors Representation for High-Fidelity Neural Implicit Dense Visual SLAM System	Neural implicit dense visual SLAM

2. Problem setting and design motivation

The specific DL-SLAM system targets dynamic environments in which a rigid dichotomy between “static” and “dynamic” observations is inadequate (Xu et al., 2 Jul 2026). Prevailing methods are described as typically discarding predefined dynamic objects, thereby ignoring that transiently static objects offer valuable geometric constraints for pose estimation. A recent alternative based on per-pixel uncertainty maps is said to recover some of this value, but to erroneously integrate transiently static objects into the static map, producing persistent artifacts. The same prior approach is also described as relying on purely geometric information, which leads to ambiguous object boundaries in the uncertainty maps.

DL-SLAM is proposed to resolve that tension rather than merely choosing one side of it. Its stated goal is to let moving objects contribute to tracking when they are temporarily static, while removing them from the final reconstructed scene when they are truly dynamic. The paper explicitly lists three contributions: a dual-level probabilistic framework that enables transiently static objects to contribute to pose estimation while preventing their incorporation into the final map; a feedback loop coupling pixel-level and object-level reasoning, where object-level pruning guides Bayesian update of per-pixel probabilities; and a dynamic-aware semantic label refinement strategy to handle occlusions and semantic inconsistency, enabling high-fidelity semantic mapping (Xu et al., 2 Jul 2026).

Within the dynamic-SLAM landscape, DL-SLAM is positioned against two families of baselines. Relative to traditional dynamic SLAM, it does not rely on a fixed set of predefined dynamic categories, although it still uses semantic understanding to delineate objects. Relative to prior 3DGS dynamic SLAM, especially WildGS-SLAM, it differs in two stated ways: it does not fuse transiently static objects into the final map, and its pixel-wise motion estimation combines semantic labels and geometry rather than using purely geometric uncertainty (Xu et al., 2 Jul 2026).

3. Gaussian scene representation and dual-level probability

DL-SLAM represents the scene as a set of 3D Gaussians

$\mathcal{G}=\{g_i\}_{i=1}^K,$

with each Gaussian $g_i$ parameterized by a mean $\mu_i \in \mathbb{R}^3$, covariance $\Sigma_i \in \mathbb{R}^{3\times3}$, opacity $o_i \in [0,1]$, RGB color $c_i \in \mathbb{R}^3$, dynamic probability $p_i \in [0,1]$, and explicit semantic label $\ell_i \in \mathbb{R}^1$ (Xu et al., 2 Jul 2026). This representation extends 3D Gaussian Splatting from appearance and geometry to dynamic and semantic state.

Rendering follows 3DGS alpha compositing: $SE(3)$0 with

$SE(3)$1

The dynamic probability map is rendered analogously,

$SE(3)$2

and semantic rendering uses Gaussian voting,

$SE(3)$3

with the label of highest cumulative weight assigned to pixel $SE(3)$4 (Xu et al., 2 Jul 2026).

The resulting model is not merely a dense radiance field with additional labels. Its key structural feature is that dynamic state is represented at two coupled levels. Pixel-level dynamic probability provides local motion evidence; object-level dynamic probability aggregates that evidence over instance-consistent Gaussian sets. This suggests that DL-SLAM treats dynamic reasoning as a scene-structure problem rather than only as an outlier-rejection problem.

4. Probabilistic estimation pipeline

DL-SLAM first constructs a temporally consistent semantic label map (Xu et al., 2 Jul 2026). The pipeline uses Recognize Anything Model (RAM) to generate open-set tags, Grounding DINO to produce class-aware bounding boxes, and MobileSAMv2 to refine them into fine-grained semantic masks. Because framewise segmentation is temporally inconsistent, the system adds cross-frame association using a fused score combining IoU and appearance similarity from CLIP; a new detection matches a tracked object if the best fused score exceeds $SE(3)$5, and unmatched detections become new object labels.

Pixel-level dynamic probability is then estimated from semantic and geometric information. For keyframes only, DL-SLAM uses ConvGRU to estimate optical flow between current keyframe $SE(3)$6 and previous keyframe $SE(3)$7. Using dense correspondences, it estimates a fundamental matrix $SE(3)$8, computes the Sampson error $SE(3)$9 for each pixel, and converts geometric inconsistency into a dynamic probability under a chi-squared model with 2 DoF: $\mathcal{G}=\{g_i\}_{i=1}^K,$0 The paper interprets larger geometric error as larger probability of motion (Xu et al., 2 Jul 2026).

DL-SLAM then enforces object coherence in 2D. For each semantic object $\mathcal{G}=\{g_i\}_{i=1}^K,$1, it aggregates the pixel probabilities inside mask $\mathcal{G}=\{g_i\}_{i=1}^K,$2 and defines the object probability $\mathcal{G}=\{g_i\}_{i=1}^K,$3 as “the average of the first quartile and the median of the sorted values.” All pixels in $\mathcal{G}=\{g_i\}_{i=1}^K,$4 are assigned that object-level 2D probability. This is designed so that significant motion dominates the object state (Xu et al., 2 Jul 2026).

The 2D probabilities are lifted to the Gaussian map. Each new Gaussian inherits dynamic probability $\mathcal{G}=\{g_i\}_{i=1}^K,$5 and semantic label $\mathcal{G}=\{g_i\}_{i=1}^K,$6 from its corresponding pixel. To mitigate temporal inertia, DL-SLAM assigns each Gaussian a recency weight

$\mathcal{G}=\{g_i\}_{i=1}^K,$7

with $\mathcal{G}=\{g_i\}_{i=1}^K,$8, and computes object-level 3D dynamic probability as

$\mathcal{G}=\{g_i\}_{i=1}^K,$9

If $g_i$0 for three consecutive keyframes, with $g_i$1, object $g_i$2 is marked dynamic and all Gaussians belonging to any object in the global dynamic object set are removed (Xu et al., 2 Jul 2026). This categorical pruning mechanism is the paper’s explicit route to an artifact-free static map.

The system closes the loop by feeding the cleaned map back into pixel-level estimation. It renders a geometrically consistent probability map $g_i$3, applies power-law scaling

$g_i$4

with $g_i$5, and performs Bayesian fusion: $g_i$6 The posterior is projected back to visible Gaussians and used to update each Gaussian’s dynamic attribute $g_i$7 (Xu et al., 2 Jul 2026). The article’s central methodological claim is therefore not only dual-level estimation, but recursive coupling between 2D evidence and 3D map state.

5. Tracking, mapping, and semantic refinement

For tracking, DL-SLAM builds on DROID-SLAM and uses a dense bundle adjustment layer (Xu et al., 2 Jul 2026). The optimization is weighted by $g_i$8, so dynamic pixels are down-weighted while stable regions, including transiently static object regions, still contribute. This placement is important: motion awareness is integrated directly into the tracking objective rather than being applied only as a post hoc mask.

This dependence on DROID-SLAM also links the method to the broader literature on deep learning-based SLAM systems. In SPAQ-DL-SLAM, DROID-SLAM is characterized as a state-of-the-art monocular visual SLAM method that combines CNN-based feature, context, and update networks with a dense bundle adjustment backend, but is also computationally heavy, with about 4.00M total parameters, 3.94M CNN parameters, and about 4.64B FLOPs in its CNN portion (Pudasaini et al., 2024). A plausible implication is that DL-SLAM inherits both the representational strength and part of the systems burden associated with a DROID-SLAM-based frontend.

For mapping, DL-SLAM incrementally grows the Gaussian map and optimizes it in a local covisible window with dynamic weighting: $g_i$9 The color loss is

$\mu_i \in \mathbb{R}^3$0

the depth loss is

$\mu_i \in \mathbb{R}^3$1

and the regularization term is

$\mu_i \in \mathbb{R}^3$2

Optimization stops when the variance of the mapping loss in the local window falls below $\mu_i \in \mathbb{R}^3$3, with $\mu_i \in \mathbb{R}^3$4, $\mu_i \in \mathbb{R}^3$5, and $\mu_i \in \mathbb{R}^3$6 (Xu et al., 2 Jul 2026).

Semantic refinement is a distinct subsystem rather than a side effect of dense rendering. To correct semantic inconsistencies, DL-SLAM matches rendered masks $\mu_i \in \mathbb{R}^3$7 with input masks $\mu_i \in \mathbb{R}^3$8 using

$\mu_i \in \mathbb{R}^3$9

If $\Sigma_i \in \mathbb{R}^{3\times3}$0, with $\Sigma_i \in \mathbb{R}^{3\times3}$1, the labels are considered potential aliases, but the merge is finalized only after the same pair matches for three consecutive keyframes. To handle fragmented regions behind dynamic objects, the system applies dynamic-aware densification using the semantic gradient

$\Sigma_i \in \mathbb{R}^{3\times3}$2

and densifies where $\Sigma_i \in \mathbb{R}^{3\times3}$3 with $\Sigma_i \in \mathbb{R}^{3\times3}$4 (Xu et al., 2 Jul 2026).

6. Experimental characterization, limitations, and significance

DL-SLAM is evaluated on TUM RGB-D dynamic, BONN, and Wild-SLAM iPhone (Xu et al., 2 Jul 2026). Tracking metrics are ATE RMSE and ATE standard deviation; rendering and reconstruction metrics are PSNR, SSIM, and LPIPS. Baselines include ORB-SLAM2, ReFusion, DynaSLAM, CFP-SLAM, Co-SLAM, NID-SLAM, RoDyn-SLAM, DynaMoN, SGS-SLAM, DG-SLAM, SDD-SLAM, and WildGS-SLAM. Reported hardware is an Intel Core i9-12900KF CPU with an NVIDIA RTX 3090 GPU.

The paper states that on TUM RGB-D, DL-SLAM outperforms all baseline methods on every sequence and achieves up to a 13% reduction in RMSE compared to the next-best method (Xu et al., 2 Jul 2026). Reported representative averages are 1.3 cm ATE RMSE on TUM versus 1.5 cm for WildGS-SLAM, and 2.2 cm on BONN versus 2.5 cm for WildGS-SLAM. For rendering, reported examples include PSNR 21.06 versus 19.69 on Wild-SLAM iPhone “street,” and 23.74 versus 22.55 on BONN “mv_box,” again compared with WildGS-SLAM.

Qualitatively, the paper emphasizes artifact-free map reconstruction, better handling of transiently static objects, and more reliable boundaries in dynamic probability maps (Xu et al., 2 Jul 2026). In the “wandering” sequence, WildGS-SLAM is described as leaving artifacts from transiently static people, whereas DL-SLAM prunes them. In “mv_box,” a box moving from table to floor is treated as dynamic and removed to avoid ghosting in either location. The semantic refinement stage is described as correcting label inconsistencies, preserving object identity over time, reconstructing occluded regions plausibly, and supporting interactive scene editing by object prompt.

The ablation study on BONN isolates the contribution of the main modules. The reported averages are: without pixel-level probability, ATE RMSE 4.8, ATE S.D. 2.8, PSNR 19.00; without object-level probability, 2.6, 1.3, 19.38; without semantic refinement, 2.4, 1.1, 18.74; and with the full DL-SLAM system, 2.2, 1.0, 19.99 (Xu et al., 2 Jul 2026). The interpretation given in the source is that pixel-level probability is critical for robust motion-aware pose estimation, object-level probability is important for categorical pruning, and semantic refinement is important for high-fidelity semantic maps.

A broader systems perspective comes from SPAQ-DL-SLAM, which treats “DL-SLAM” as a class of deep learning-based visual SLAM methods and studies structured pruning and 8-bit post-training static quantization for embedded deployment (Pudasaini et al., 2024). That work reports that moderate compression can reduce FLOPs and model size while sometimes improving ATE and AUC, but that performance can vary on high-angular-velocity EuRoC Vicon Room sequences. This suggests that the significance of DL-SLAM now extends beyond algorithm design in dynamic scenes to questions of resource efficiency, deployment regime, and environment-specific co-design.