MyGO-Splat: RGB-Only Gaussian SLAM

• MyGO-Splat is an RGB-only Gaussian SLAM framework that uses active Gaussian maps to counteract scale ambiguity and depth inconsistency.
• It analytically rasterizes Gaussian primitives into pixel-wise depth and normals and aligns monocular depth priors via scale-aware adaptive optimization.
• Evaluations on Replica, TUM RGB-D, and ScanNet demonstrate near RGB-D performance with low ATE and high rendering quality through closed-loop feedback.

MyGO-Splat is an RGB-only Gaussian SLAM framework that targets two persistent monocular failure modes: scale ambiguity and the absence of geometric self-correction. Its central design choice is to make the Gaussian map an active supervisory signal rather than a passive rendering structure. To do so, it analytically rasterizes Gaussian primitives into pixel-wise depth and surface normals, aligns monocular depth priors from a foundation model to the globally optimized Gaussian map, and feeds the resulting geometry back into tracking and mapping in a closed loop. The system combines flow-based online monocular tracking, loop closure and global bundle adjustment, analytical Gaussian depth/normal rendering, scale-aware adaptive alignment, and joint appearance-geometry optimization, with evaluations on Replica, TUM RGB-D, and ScanNet reporting RGB-only performance close to RGB-D Gaussian SLAM baselines (Zhu et al., 29 Jun 2026).

1. Problem definition and scene representation

MyGO-Splat is formulated for monocular dense SLAM with a 3D Gaussian map. The paper frames prior RGB-only Gaussian SLAM systems as effectively open-loop: pseudo-depth or monocular priors may help mapping, but refined geometry does not actively regulate later pose estimation, so scale drift and depth inconsistency remain difficult to suppress (Zhu et al., 29 Jun 2026).

The map is represented as a set of anisotropic Gaussian primitives. The paper writes each primitive as

$$\mathcal{G}_i=\delta_i \mathcal{N}(\mathbf{P}_i,\mathbf{\Sigma}_i),$$

where $\mathbf{P}_i$ is the Gaussian center, $\mathbf{\Sigma}_i$ is the covariance, and $\delta_i \in [0,1]$ is the opacity; color is represented by third-order spherical harmonics (Zhu et al., 29 Jun 2026). The text states that the covariance is composed from scale $\mathbf{S}$ and rotation $\mathbf{R}$, but it does not print the explicit factorization. This suggests continuity with standard 3DGS parameterization, while leaving the exact decomposition to inherited implementation details.

The paper’s conceptual shift is that the Gaussian map is not only a radiance representation. It is also a geometric reference from which depth and surface orientation can be analytically rendered and then reintroduced into SLAM optimization. A plausible implication is that the map begins to function as an internal geometric sensor, not merely as a photometric reconstruction.

2. Closed-loop SLAM architecture

The input is a monocular RGB stream $\{\mathbf{I}_i\}_{i=1}^N$, and the outputs are the camera trajectory $\{\mathbf{T}_i\}_{i=1}^N$, the Gaussian scene representation $\mathcal{G}$, rendered RGB/depth/normal maps, and mesh extraction from the Gaussian map (Zhu et al., 29 Jun 2026). The frontend is keyframe-based and follows DROID-SLAM-style recurrent optical flow. A graph $(\mathcal{V},\mathcal{E})$ is maintained, whose nodes store keyframe poses $\mathbf{P}_i$0 and depth maps $\mathbf{P}_i$1, while edges encode dense correspondences. A new frame becomes a keyframe when the average flow distance to the previous keyframe exceeds a threshold $\mathbf{P}_i$2. After collecting $\mathbf{P}_i$3 initial keyframes, the system performs initial bundle adjustment, then continues with dense local BA in a sliding window (Zhu et al., 29 Jun 2026).

The local BA objective is printed in the paper with typographical corruption, but its intended structure is clear: it minimizes a confidence-weighted discrepancy between predicted target points $\mathbf{P}_i$4 from the flow network and the reprojection operator $\mathbf{P}_i$5, with $\mathbf{P}_i$6 acting as a correspondence confidence term. Optimization is performed with damped Gauss-Newton (Zhu et al., 29 Jun 2026).

Long-range consistency is handled by loop closure and global BA. Loop candidates are retrieved with global image/geometric descriptors from EigenPlaces using FAISS nearest-neighbor search, after which a global factor graph is optimized over all keyframe poses and dense depth maps (Zhu et al., 29 Jun 2026). This global optimization defines the coordinate frame used later for scale-aware depth alignment.

The closed loop is completed by a separate geometric prior branch. MyGO-Splat uses a 3D vision foundation model, specifically VGGT, but retains only its depth information. Those monocular priors are aligned to the Gaussian-map geometry, used to optimize the map, and the refined Gaussian-rendered depth is periodically fed back to the BA/tracking layer. The paper explicitly frames this as replacing unstable per-frame pseudo-depth with multi-view-consistent Gaussian-rendered depth (Zhu et al., 29 Jun 2026).

3. Analytical depth and normal rasterization

The core geometric novelty is analytical rasterization of Gaussian depth and surface normals. Instead of assigning a pixel depth from the Gaussian center, MyGO-Splat uses a local affine model in a transformed ray space where the set of Gaussian-ray intersection maxima becomes coplanar. This yields a per-Gaussian planar depth field over the image support (Zhu et al., 29 Jun 2026).

The paper gives the pixel depth as

$\mathbf{P}_i$7

where $\mathbf{P}_i$8 is the depth of the Gaussian center, $\mathbf{P}_i$9 is the image coordinate of that center, and $\mathbf{\Sigma}_i$0 is determined by Gaussian covariance and camera intrinsics (Zhu et al., 29 Jun 2026). This depth is therefore spatially varying across the projected support of the Gaussian rather than piecewise constant.

For normals, the paper derives a plane normal in ray space,

$\mathbf{\Sigma}_i$1

and then maps it to camera coordinates through

$\mathbf{\Sigma}_i$2

The printed notation for $\mathbf{\Sigma}_i$3, $\mathbf{\Sigma}_i$4, and $\mathbf{\Sigma}_i$5 is not fully expanded in the paper, but their roles are explicit: $\mathbf{\Sigma}_i$6 encodes local depth variation, $\mathbf{\Sigma}_i$7 is associated with the local affine/ray-space transform, and $\mathbf{\Sigma}_i$8 is the Jacobian used to transform normals back to the camera frame (Zhu et al., 29 Jun 2026).

Color and depth are composited front-to-back: $\mathbf{\Sigma}_i$9 where $\delta_i \in [0,1]$0 is the Gaussian color contribution and $\delta_i \in [0,1]$1 its per-pixel translucency (Zhu et al., 29 Jun 2026). The paper does not provide a separate full normal-compositing equation, but it does define analytical per-Gaussian normals and uses them in the mapping loss. This suggests that MyGO-Splat’s geometric supervision is tied directly to Gaussian shape rather than to an auxiliary surface estimator.

4. Scale-aware alignment and multi-objective map optimization

MyGO-Splat introduces an uncertainty-weighted affine depth alignment between rendered Gaussian depth $\delta_i \in [0,1]$2 and monocular prior depth $\delta_i \in [0,1]$3: $\delta_i \in [0,1]$4 Here $\delta_i \in [0,1]$5 is the set of high-confidence tracked pixels, $\delta_i \in [0,1]$6 is a confidence weight derived from photometric residuals and depth stability, and $\delta_i \in [0,1]$7 are per-keyframe scale and shift parameters (Zhu et al., 29 Jun 2026). The aligned proxy depth is then used for Gaussian-map supervision. This is the paper’s main mechanism for bringing monocular priors into the metric frame maintained by global BA.

The mapping stage follows standard 3DGS defaults with Mip-Splatting-style filtering, GOF-based densification, and Adam optimization, but adds a geometric-enhanced multi-objective loss. The individual terms are an RGB reconstruction loss, a depth consistency loss against the aligned proxy depth, a discrepancy regularizer over Gaussians along the same ray, and a normal-consistency loss between Gaussian analytical normals and normals computed from prior depth (Zhu et al., 29 Jun 2026).

The paper defines the loss components in prose and notation as follows:

• $\delta_i \in [0,1]$8: reconstruction of observed image $\delta_i \in [0,1]$9 by rendered image $\mathbf{S}$0
• $\mathbf{S}$1: reconstruction of aligned depth $\mathbf{S}$2 by rendered depth $\mathbf{S}$3
• $\mathbf{S}$4: discrepancy penalty $\mathbf{S}$5 for Gaussians along the same ray
• $\mathbf{S}$6: normal consistency $\mathbf{S}$7

The total objective uses

$\mathbf{S}$8

with $\mathbf{S}$9 on ScanNet (Zhu et al., 29 Jun 2026). The large weight on $\mathbf{R}$0 reflects the paper’s concern with floaters and multi-surface inconsistency along a ray. A plausible implication is that this term functions as a surface-coherence prior adapted to Gaussian rasterization rather than to meshes or TSDFs.

5. Empirical performance and ablations

The evaluation uses Replica, TUM RGB-D, and ScanNet, following NICE-SLAM sequence selection where applicable, with results averaged over five runs. Metrics are ATE RMSE for pose, PSNR/SSIM/LPIPS for rendering, and Accuracy/Completion for geometry on Replica (Zhu et al., 29 Jun 2026).

Dataset	MyGO-Splat results	Reported context
Replica	ATE 0.26 cm, PSNR 38.33 dB, SSIM 0.972, LPIPS 0.032, Acc. 1.54 cm, Comp. 3.69 cm, $\mathbf{R}$1 FPS	Best RGB-only result in most categories; close to RGB-D baselines
TUM RGB-D	ATE 1.14 cm, PSNR 25.01 dB, SSIM 0.813, LPIPS 0.125	Best among RGB-only methods; close to RTG-SLAM
ScanNet	ATE 7.22 cm, PSNR 29.09 dB, SSIM 0.878, LPIPS 0.232	Best ATE and SSIM among RGB-only methods

On Replica, the method is compared with Photo-SLAM, MonoGS, Splat-SLAM, and SEGS-SLAM on the RGB-only side, and with SplaTAM, RTG-SLAM, and GS-ICP SLAM on the RGB-D side. The paper states that MyGO-Splat is the best RGB-only system in most Replica metrics and remains close to GS-ICP SLAM and RTG-SLAM despite using no depth sensor (Zhu et al., 29 Jun 2026). On TUM RGB-D, it slightly trails RTG-SLAM in ATE but surpasses the other RGB-only baselines. On ScanNet, it reports the best ATE and SSIM among RGB-only methods, while the paper notes that global monocular scale remains challenging on this dataset (Zhu et al., 29 Jun 2026).

The ablation study isolates loop closure, closed-loop geometric feedback, and geometric-enhanced multi-objective optimization. Removing loop closure increases ATE from 0.26 cm to 0.45 cm on Replica, with smaller effects on rendering. Removing closed-loop geometric feedback degrades ATE to 0.58 cm, PSNR to 35.12, Accuracy to 3.86 cm, and Completion to 5.24 cm. Removing geometric-enhanced multi-objective optimization leaves ATE near 0.29 cm and slightly increases PSNR to 38.45, but geometry deteriorates sharply to Accuracy 5.42 cm and Completion 9.85 cm (Zhu et al., 29 Jun 2026). This pattern supports the paper’s distinction between trajectory stabilization and geometric quality: the feedback loop primarily stabilizes scale and pose, while the geometric losses primarily improve structural fidelity.

6. Position within Gaussian-SLAM research and limitations

Within the broader Gaussian robotics literature, MyGO-Splat occupies the monocular closed-loop end of the design space. Related systems use Gaussian maps for other roles: object-aware SLAM and open-vocabulary querying in Go-SLAM (Pham et al., 2024), active exploration in RT-GuIDE (Tao et al., 2024), language-grounded multi-goal navigation memory in LagMemo (Zhou et al., 28 Oct 2025), and LiDAR-native odometry and mapping in Splat-LOAM (Giacomini et al., 21 Mar 2025). This suggests that 3DGS has become a general robotics substrate spanning reconstruction, semantics, exploration, and navigation, while MyGO-Splat specifically focuses on geometry-driven self-correction for RGB-only SLAM.

The paper’s limitations are partly explicit and partly structural. It assumes static scenes and a keyframe-based monocular pipeline; no dedicated dynamic-scene mechanism is described (Zhu et al., 29 Jun 2026). It depends on a foundation-model depth prior, specifically VGGT, whose scale and stability errors motivate the alignment stage. The ScanNet results indicate that global monocular scale remains difficult in challenging scenes. Several implementation details are not fully specified in the text: the explicit covariance decomposition, the closed-form expression for $\mathbf{R}$2, the precise meaning of $\mathbf{R}$3, the full Jacobian $\mathbf{R}$4, and detailed Gaussian spawn/prune criteria. Runtime is reported only coarsely—generally $\mathbf{R}$5 FPS, with an outdoor demonstration at 5 FPS—and no memory footprint is given, although experiments were run on an NVIDIA A100 with 40 GB VRAM (Zhu et al., 29 Jun 2026).

A plausible interpretation is that MyGO-Splat’s main contribution is not a new Gaussian primitive or a new tracker in isolation, but a systems-level closure of the monocular SLAM loop: depth priors shape the Gaussian map, the Gaussian map analytically renders geometry, and that rendered geometry in turn regulates the depth variables and camera poses that produced it. In that sense, MyGO-Splat redefines the Gaussian map from an endpoint of reconstruction into an internal geometric feedback mechanism for monocular SLAM (Zhu et al., 29 Jun 2026).