MipSLAM: Multipath, Vision & Semantic SLAM

• MipSLAM is a term that encompasses diverse SLAM frameworks, including radio multipath SLAM, alias-free Gaussian splatting SLAM, and multi-modal semantic SLAM.
• The radio approach uses Bayesian inference to treat multipath signals as structured geometric information for joint trajectory and environment estimation.
• The vision and semantic variants improve pose estimation and mapping fidelity using frequency-aware anti-aliasing and fusion of visual and LiDAR data in dynamic scenes.

MipSLAM denotes distinct SLAM formulations in recent arXiv literature rather than a single canonical algorithm. In wireless localization, the term is used for multipath-based simultaneous localization and mapping, a class of Bayesian methods that treats reflected radio paths as geometric information for joint trajectory estimation and environment reconstruction; this line includes feature-based, direct-from-waveform, cooperative, and coherent D-MIMO/XL-MIMO variants (Li et al., 2024, Liang et al., 2023, Deutschmann et al., 21 Apr 2026). In computer vision, “MipSLAM: Alias-Free Gaussian Splatting SLAM” names a frequency-aware 3D Gaussian Splatting framework for anti-aliased novel view synthesis and robust pose estimation under varying camera configurations (Li et al., 7 Mar 2026). A separate dynamic-scene robotics system in the same name neighborhood is MMS_SLAM, a multi-modal semantic SLAM framework that combines RGB instance segmentation with LiDAR clustering (Wang et al., 2022).

1. Terminological scope

In the wireless literature, MipSLAM is used interchangeably with multipath-based SLAM or MP-SLAM. Its central premise is that line-of-sight can be blocked or unreliable, while reflected single-bounce paths still encode strong geometric structure. Instead of treating multipath as interference, the environment is modeled so that multipath components can be jointly used to estimate the moving terminal trajectory and reconstruct a map of reflective surfaces. This usage spans single-base-station mmWave MIMO, multi-base-station MIMO, cooperative multi-terminal settings, direct inference from raw RF waveforms, and coherent distributed-array processing (Liang et al., 2024, Leitinger et al., 2024).

A distinct usage appears in the vision paper “MipSLAM: Alias-Free Gaussian Splatting SLAM,” where MipSLAM is a 3DGS-based SLAM framework whose defining properties are frequency-aware anti-aliasing, spectral-aware pose graph optimization, and a local frequency-domain perceptual loss (Li et al., 7 Mar 2026). This formulation belongs to the Gaussian-splatting SLAM family rather than the radio-SLAM lineage.

A further source of ambiguity is MMS_SLAM, which is a multi-modal semantic SLAM system for complex dynamic environments. MMS_SLAM combines instance segmentation, multi-modal fusion, localization, and global optimization and mapping, but it is not a multipath-radio method and not a Gaussian-splatting method (Wang et al., 2022).

A recurring misconception is to treat these papers as a single methodological lineage. The cited works instead show that “MipSLAM” functions as an overloaded label across at least three technically distinct research directions: radio multipath SLAM, alias-free Gaussian-splatting SLAM, and multi-modal semantic dynamic-scene SLAM.

2. Radio-frequency MipSLAM: multipath as geometric signal

Radio-frequency MipSLAM is formulated as a Bayesian SLAM problem in which RF measurements are linked to latent map features and to the mobile state. A standard abstraction is the virtual anchor: for a single-bounce specular reflection, the reflective surface is represented by the mirror image of a physical anchor or base station across that surface. The same literature also generalizes beyond virtual anchors to point scatterers and to surface feature vectors, depending on the propagation model and the sensing infrastructure (Leitinger et al., 2024, Li et al., 2024, Deutschmann et al., 21 Apr 2026).

In direct waveform formulations, the received signal at anchor or panel $j$ is modeled as a superposition of feature contributions plus noise,

$$\mathbf{z}_k^{(j)} = \sum_{n=1}^{N_k^{(j)}} r_{k,n}^{(j)} \rho_{k,n}^{(j)} \mathbf{h}_{k,n}^{(j)} + \boldsymbol{\epsilon}_k^{(j)},$$

with binary existence variables $r_{k,n}^{(j)}$, complex amplitudes $\rho_{k,n}^{(j)}$, feature-dependent delayed waveform samples $\mathbf{h}_{k,n}^{(j)}$, and additive residual noise $\boldsymbol{\epsilon}_k^{(j)}$. Conditioned on the agent state, feature states, and noise variance, the measurement is modeled as complex Gaussian with covariance

$$\mathbf{C}_k^{(j)} = \eta_k^{(j)} \mathbf{I}_M + \sum_{n=1}^{N_k^{(j)}} r_{k,n}^{(j)} \gamma_{k,n}^{(j)} \mathbf{h}_{k,n}^{(j)} \mathbf{h}_{k,n}^{(j)H},$$

which makes the waveform a coherent superposition of all existing potential features (Liang et al., 2024).

Feature-based mmWave MIMO variants operate on extracted channel parameters such as delay, angle of arrival, angle of departure, and amplitude. The corresponding Bayesian posterior jointly covers the mobile state, map-feature states, existence variables, and association variables. A factor graph then expresses dynamic priors, measurement likelihoods, feature births and deaths, and data-association constraints, while belief propagation approximates marginal posteriors for localization and mapping (Li et al., 2024).

This framework is significant because it reinterprets nuisance propagation effects as structured observations of the environment. In single-BS scenarios, multipath is essential rather than optional; in multi-BS or distributed-array settings, it becomes a means to combine aperture, geometry, and propagation structure within one SLAM posterior.

3. Direct and coherent radio MipSLAM

Direct multipath-based SLAM removes the conventional preprocessing stage in which a channel estimator first extracts multipath components from the received waveform. The argument is that finite bandwidth and limited resolution can merge closely spaced paths, causing missed detections, false alarms, or incorrect parameter extraction before SLAM inference even begins. Direct-SLAM therefore uses the raw received radio signal vector as the measurement and represents the full statistical data-generation process by a factor graph (Liang et al., 2023, Liang et al., 2024).

The inference engine in these direct methods is sequential belief propagation with particle-based beliefs for the agent, feature states, and noise variables. Exact update messages are intractable because the waveform likelihood couples all latent features, so the method replaces exact mixtures by Gaussian messages via moment matching. The resulting complexity scales linearly with the number of potential features in the covariance calculation, while retaining the ability to exploit waveform information that would be discarded by channel estimation. In synthetic and real SISO scenarios, the direct method outperforms channel-estimation-based baselines in localization and mapping accuracy, especially when several virtual anchors have similar delays and the channel estimator cannot resolve multiple paths (Liang et al., 2024).

The 2026 coherent extension generalizes direct MP-SLAM to synchronized D-MIMO and XL-MIMO infrastructures. Its core innovation is a nonzero-mean Type-II likelihood,

$$f(\mathbf z_n^{(j)}\mid x_n,y_n^{(j)},r_n^{(j)},\eta_n^{(j)}) = \mathcal{CN}\!\left(\mathbf z_n^{(j)};\,\boldsymbol\mu_n^{(j)},\mathbf C_n^{(j)}\right),$$

with

$$\boldsymbol\mu_n^{(j)}=\sum_{s\in\mathcal S_n} r_{s,n}^{(j)}\,\mu_{s,n}^{(j)}\,\bm\psi(x_n,p_{s,n}),$$

and

$$\mathbf C_n^{(j)}=\eta_n^{(j)}\mathbf I +\sum_{s\in\mathcal S_n} r_{s,n}^{(j)}\,\sigma_{s,n}^{(j)}\,\bm\psi(x_n,p_{s,n})\bm\psi(x_n,p_{s,n})^H.$$

Here the mean carries the coherent complex field shared across synchronized PAs or subarrays, while the covariance models the remaining signal power and acts as a noncoherent energy term (Deutschmann et al., 21 Apr 2026).

This coherent formulation also replaces PA-local virtual-anchor duplication with a surface feature vector model in which each physical reflecting surface corresponds to a global feature position $$\mathbf{z}_k^{(j)} = \sum_{n=1}^{N_k^{(j)}} r_{k,n}^{(j)} \rho_{k,n}^{(j)} \mathbf{h}_{k,n}^{(j)} + \boldsymbol{\epsilon}_k^{(j)},$$0. The SFV model supports near-field propagation through geometry-dependent steering vectors, visibility effects and partial obstruction, and a two-layer existence model comprising a global feature-existence variable and PA-specific path-existence variables. Inference is implemented as particle-based Bayesian belief propagation on a factor graph with moment-matched Gaussian approximations, and a GPU-parallel implementation yields an order-of-magnitude speedup over the earlier CPU implementation. Simulation results show gains over noncoherent methods and approach the posterior Cramér–Rao lower bound, while the principal requirement is phase coherence across the infrastructure (Deutschmann et al., 21 Apr 2026).

4. Generalizations of radio MipSLAM

The radio-MipSLAM literature extends the classical virtual-anchor model in several orthogonal directions. One direction is heterogeneity of map features. Another is non-ideal surface modeling. A third is cooperation among multiple mobile terminals. Together these papers broaden MP-SLAM from a single-agent, single-measurement-per-surface, ideal-specular setting to a more general inference framework.

Paper	Generalization	Stated effect
(Li et al., 2024)	Flat-surface specular reflectors and point scatterers; amplitude information	Reliable detection of weak map features associated with MPCs of very low SNRs
(Wielandner et al., 2024)	Per-feature delay and angle dispersion; multiple-measurement-to-feature association	Joint estimation of the positions and dispersion extents of ideal and non-ideal reflective surfaces
(Leitinger et al., 2024)	Multiple MTs, map fusion, cooperative localization, IMU	More robust mapping and higher localization accuracy

In the heterogeneous-feature formulation, flat-surface specular reflections are represented through virtual-anchor geometry, while point scatterers are represented as discrete spatial points. The Bayesian model supports sequential detection and estimation of interacting model parameters, map-feature states, and the mobile state, and amplitude enters the likelihood to improve detection of weak features that geometry alone might miss. The experimental setting uses real mmWave MIMO measurements with a single base station in a dynamic outdoor scenario (Li et al., 2024).

In the non-ideal-surface formulation, the conventional assumption “one VA $$\mathbf{z}_k^{(j)} = \sum_{n=1}^{N_k^{(j)}} r_{k,n}^{(j)} \rho_{k,n}^{(j)} \mathbf{h}_{k,n}^{(j)} + \boldsymbol{\epsilon}_k^{(j)},$$1 one MPC” is explicitly relaxed. A single physical surface may generate multiple resolved MPCs because of delay and angular dispersion. Each potential virtual anchor therefore carries dispersion parameters $$\mathbf{z}_k^{(j)} = \sum_{n=1}^{N_k^{(j)}} r_{k,n}^{(j)} \rho_{k,n}^{(j)} \mathbf{h}_{k,n}^{(j)} + \boldsymbol{\epsilon}_k^{(j)},$$2, and the algorithm uses multiple-measurement-to-feature association inside a Bayesian particle-based sum-product algorithm. For ideal reflectors, the dispersion extents can collapse toward zero; for rough or non-ideal reflectors, the same feature can explain a cluster of measurements (Wielandner et al., 2024).

The cooperative extension introduces multiple mobile terminals that exchange information and make MT-to-MT LOS measurements in addition to BS-to-MT multipath observations. The same VA may then be observed by different MTs at different times, and the algorithm fuses these observations into a single estimate while also supporting cooperative localization and IMU-assisted navigation. In the reported simulations, the strongest configuration combines PVA fusion, cooperation, IMU, and MIMO; the worst case is the state-of-the-art MP-SLAM setting without PVA fusion and cooperation, even if IMU and MIMO are present (Leitinger et al., 2024).

5. MipSLAM as alias-free Gaussian Splatting SLAM

In the vision literature, MipSLAM is a frequency-aware 3D Gaussian Splatting SLAM framework. Its purpose is to maintain rendering fidelity and pose accuracy under camera parameter changes such as resolution, focal length, and zoom. The system has three modules: tracking, alias-free Gaussian mapping, and Spectral-Aware Pose Graph Optimization. Its central claim is that aliasing in 3DGS-SLAM is not only a rendering issue; it also contaminates the gradients used in pose optimization (Li et al., 7 Mar 2026).

The rendering model starts from the standard 3DGS alpha-compositing equation,

$$\mathbf{z}_k^{(j)} = \sum_{n=1}^{N_k^{(j)}} r_{k,n}^{(j)} \rho_{k,n}^{(j)} \mathbf{h}_{k,n}^{(j)} + \boldsymbol{\epsilon}_k^{(j)},$$3

with an analogous expression for depth. Standard 3DGS evaluates $$\mathbf{z}_k^{(j)} = \sum_{n=1}^{N_k^{(j)}} r_{k,n}^{(j)} \rho_{k,n}^{(j)} \mathbf{h}_{k,n}^{(j)} + \boldsymbol{\epsilon}_k^{(j)},$$4 at the pixel center. MipSLAM replaces this point evaluation with a pixel-footprint integral and approximates the contribution numerically by Elliptical Adaptive Anti-aliasing. The projected Gaussian covariance is decomposed as $$\mathbf{z}_k^{(j)} = \sum_{n=1}^{N_k^{(j)}} r_{k,n}^{(j)} \rho_{k,n}^{(j)} \mathbf{h}_{k,n}^{(j)} + \boldsymbol{\epsilon}_k^{(j)},$$5, sampling is performed in the principal-axis frame, and the importance weighting depends on anisotropy and on boundary proximity. The stated objective is close-to-analytic fidelity without the memory and computational cost of full analytic integration (Li et al., 7 Mar 2026).

The back-end replaces purely spatial pose-graph reasoning with spectral-spatial optimization. SA-PGO represents poses as a trajectory signal, computes sliding-window DFTs, derives spectral signatures from the power-spectrum centroid, measures spectral coherence between poses, forms a weighted adjacency matrix, and analyzes the normalized graph Laplacian. The Fiedler value is used as a measure of algebraic connectivity, and spectral confidence weights influence edge selection and optimization strength. A local frequency-domain perceptual loss complements this by comparing magnitude and phase spectra of rendered and ground-truth depth patches, with higher weight on spectrally complex patches (Li et al., 7 Mar 2026).

The reported evaluations use Replica and TUM RGB-D at multiple simulated resolutions. On Replica, the average over 8 sequences is PSNR 36.19, SSIM 0.963, and LPIPS 0.043; on TUM, the average over 5 tested scales is PSNR 22.946, SSIM 0.814, and LPIPS 0.216. For localization on Replica, the average ATE RMSE is 0.22 cm, compared with 0.37 cm for the variant with MipSplatting-style anti-aliasing, 0.31 cm for Gaussian-SLAM, 0.29 cm for LoopySLAM, 0.58 cm for MonoGS, and 0.38 cm for SplaTAM. The paper also reports a time breakdown on Replica Room0 at $$\mathbf{z}_k^{(j)} = \sum_{n=1}^{N_k^{(j)}} r_{k,n}^{(j)} \rho_{k,n}^{(j)} \mathbf{h}_{k,n}^{(j)} + \boldsymbol{\epsilon}_k^{(j)},$$6: 2.17 ms for $$\mathbf{z}_k^{(j)} = \sum_{n=1}^{N_k^{(j)}} r_{k,n}^{(j)} \rho_{k,n}^{(j)} \mathbf{h}_{k,n}^{(j)} + \boldsymbol{\epsilon}_k^{(j)},$$7-blending, 0.71 ms for the FPS-related step, 868 ms for tracking, 187 ms for mapping, and 65 ms for SA-PGO. The stated limitations are computational overhead, sensitivity to graph quality, the approximation nature of EAA, and uncertainty about generalization to outdoor, highly dynamic, or low-texture scenes (Li et al., 7 Mar 2026).

6. MMS_SLAM and dynamic-scene semantic SLAM

MMS_SLAM is a separate system whose relevance to the MipSLAM name arises from query ambiguity rather than methodological identity. It is a multi-modal semantic SLAM system designed for complex dynamic environments, especially settings in which humans, AGVs, and other moving objects violate the static-scene assumption of conventional SLAM. The architecture has four modules: instance segmentation / semantic learning, multi-modal fusion, localization, and global optimization and mapping (Wang et al., 2022).

The front-end uses a lightweight SOLOv2-style instance segmentation model and improves its backbone with Recursive Feature Pyramid and Switchable Atrous Convolution. RFP is described as a mechanism that lets the network “look twice or more” by feeding FPN outputs back into the bottom-up backbone layers, while SAC adaptively combines outputs from different atrous rates. The semantic output is projected into the LiDAR frame and fused with Euclidean clustering. Morphological dilation expands dynamic-mask boundaries to compensate for motion blur, and the geometric correction rule classifies a point cluster as dynamic when 90% of its points are dynamic-labeled. Static points are then used for LiDAR odometry and localization, while both static and dynamic points are retained for map construction (Wang et al., 2022).

Localization is based on edge and planar features and minimizes point-to-edge and point-to-plane residuals by the Gauss-Newton method. The system updates local feature maps continuously, updates the global dense map only on keyframes, reprojects 3D points into the image plane for coloring, and voxel-downsamples the map after each update to prevent memory overflow. The dynamic map is kept separately for visualization or downstream tasks such as motion planning (Wang et al., 2022).

The reported segmentation results on COCO are: SOLOv2 with loss 0.52, mAP 38.8%, and inference 54 ms; SOLOv2 + RFP with loss 0.36, mAP 41.2%, and inference 64 ms; SOLOv2 + SAC with loss 0.39, mAP 39.8%, and inference 59 ms; and SOLOv2 + DetectoRS (Ours) with loss 0.29, mAP 43.4%, and inference 71 ms. The combined RFP+SAC system improves mAP by 5.9% over baseline SOLOv2 at a cost of about 17 ms extra inference time. The localization comparison reports ATDE 4.834 cm and MTDE 1.877 cm without semantic recognition, ATDE 1.273 cm and MTDE 0.667 cm with vision-based semantic recognition, and ATDE 0.875 cm and MTDE 0.502 cm for the multi-modal method. The abstract states that the framework can build a static dense map at more than 10 Hz, and the underlying localization component can run at 30 Hz. The paper also states that both the training data and the proposed method are open sourced (Wang et al., 2022).

Across these usages, the common theme is not a shared implementation but a shared SLAM ambition: to preserve information that conventional pipelines discard. In radio MipSLAM that information is multipath structure and, in the coherent case, infrastructure-wide phase relations; in Gaussian-splatting MipSLAM it is the frequency content lost by point-sampled splatting and purely spatial pose optimization; in MMS_SLAM it is dynamic-scene structure that single-modality segmentation leaves unresolved.