Bidirectionally Coupled SLAM

• Bidirectionally coupled SLAM is a framework where at least two subsystems mutually refine their estimates through cyclic information exchange, improving overall mapping performance.
• It integrates diverse modalities such as cameras, LiDAR, and IMU using joint optimization methods like bundle adjustment, EM, or bilevel optimization to concurrently solve multiple estimation tasks.
• The approach enhances accuracy and robustness in dynamic and multimodal environments, with applications in autonomous driving, robotics, and augmented reality despite higher computational complexity.

A bidirectionally coupled SLAM (Simultaneous Localization and Mapping) framework denotes a SLAM system in which at least two subsystems (modalities, algorithmic layers, or sensor streams) exchange informative signals in both directions such that each mutually corrects, constrains, or regularizes the other during operation or optimization. This paradigm generalizes classic "tightly coupled" SLAM—which typically fuses modalities or tasks in a monolithic cost function—by making explicit the cyclic, information-exchange pathways and the iterative, alternating refinement, leading to enhanced accuracy, robustness, and real-time performance in challenging, dynamic, or multimodal environments.

1. Conceptual Foundations and Theoretical Formulations

Bidirectionally coupled SLAM systems formalize the recursive, mutual interaction between at least two tracks of estimation: for example, camera and LiDAR odometry, static scene mapping and dynamic object tracking, visual odometry and dense Gaussian scene reconstruction, or segmentation and localization. The core theoretical ingredient is a coupled optimization, typically expressed as a single joint (robustified) cost function whose solution strategy alternates variable updates across the bidirectionally coupled modules.

A canonical example is DynaSLAM II, which models the static scene and moving object structure, as well as camera and dynamic object trajectories, through a single bundle adjustment (BA) objective:

$$J(\theta) = \sum_{k=1}^{N_c} \left[ \sum_{i\in S_k} \rho\left(\frac{\|u_{ik} - \pi(T_{C_k}^{-1} X_i)\|^2}{\sigma_{ik}^2}\right) + \sum_{j\in O_k} \left( \rho(\cdot)_{\text{dyn}} + \rho(\cdot)_{\text{vel}} + \sum_\ell \rho(\cdot)_{\text{dyn\_xyz}} \right) \right]$$

where $\theta$ aggregates camera, static map, dynamic object, and dynamic landmark variables. This factor graph framework induces bidirectional dependencies: static SLAM informs dynamic tracking, and dynamic object structure feeds back to improve camera localization and local scale/orientation (Bescos et al., 2020).

In learning-based frameworks, such as iSLAM, the bidirectional coupling is formalized via bilevel or alternating optimization. iSLAM expresses this as: $$\min_\theta\; \mathcal L(f_\theta, \mathbf P^*(\theta), \mathbf M) \quad \mathrm{s.t.}\; \mathbf P^*(\theta) = \arg\min_{\mathbf P} \mathcal L(f_\theta, \mathbf P, \mathbf M)$$ which enables global residuals from the pose graph optimization to improve front-end odometry weights via gradient backpropagation (Fu et al., 2023).

Expectation–Maximization (EM) is leveraged for dense, direct SLAM integration (GSO-SLAM), where the pose/depth (VO) and Gaussian scene (GS) representations are co-refined, each step using outputs of the other as priors or pseudo-observations (Yeon et al., 12 Feb 2026).

2. Architecture: Information Flow and Modalities

Bidirectionally coupled SLAM frameworks can be categorized by the modalities or tasks that are fused, and by the scheme used for information exchange.

• Static/dynamic scene coupling: DynaSLAM II jointly optimizes static environment points and dynamic object tracks, using semantic segmentation for initial association, and fusing object-induced parallax to stabilize camera orientation and scale (Bescos et al., 2020).
• Multimodal sensor coupling: VIRAL SLAM integrates camera, IMU, lidar, and UWB, with separate local (sliding window) and global (bundle adjustment) graphs. Back-propagation of updated UWB extrinsics to the local thread closes the bidirectional loop (Nguyen et al., 2021).
• Dense geometry–tracking coupling: GSO-SLAM's EM scheme iterates between direct sparse odometry (tracking window BA) and a 2D Gaussian splatting map, with depth and pose pseudo-observations traversing in both directions each iteration (Yeon et al., 12 Feb 2026).
• Segmentation/localization co-optimization: Mutual gating between instance segmentation and SLAM, where segmentation prunes dynamic features to stabilize SLAM, and SLAM pose refines and coherently warps segmentation masks (Wang et al., 2018).
• LiDAR-visual geometric features: Tightly-coupled visual–LiDAR SLAM maps visual line features to LiDAR 3D lines, enforcing correspondence in both BA and ICP, with direction information being exchanged in both directions (Cao et al., 2023).

Information exchange often occurs both at the data-association/front-end (labeling of features, initialization of constraints) and in the back-end (cost function or residuals involving variables from both sides), with factor graphs being the unifying language for implementation.

3. Algorithms and Optimization Strategies

Bidirectionally coupled SLAM frameworks commonly implement an alternating or joint optimization strategy, with the coupling enforced at the level of residuals, priors, or initialization.

Bundle Adjustment with Dynamic Objects:

DynaSLAM II employs a block-sparse Gauss–Newton BA, with factors linking camera poses, static points, dynamic object poses, dynamic landmarks, and object velocities. Schur complement marginalization is used to decouple variables, yielding a small linear system for the coupled motion states (Bescos et al., 2020).

Sliding Window and Global Graph Fusion:

VIRAL SLAM uses decoupled local and global optimization, re-enabling extrinsic UWB residuals post-global BA, with downstream corrections of local state estimates by the updated calibration (Nguyen et al., 2021).

Alternating EM Optimization:

GSO-SLAM alternates E-step (scene Gaussian update with fixed pose/depth) and M-step (pose/depth update with fixed scene), with depth maps and rendered depth images exchanged between the two modules. This scheme achieves full real-time throughput with no additional overhead (Yeon et al., 12 Feb 2026).

Bilevel/Bi-variational Learning:

iSLAM employs bilevel optimization where the front-end neural odometry model is updated via global pose-graph losses, propagating long-range geometric corrections back to the local model (Fu et al., 2023).

Segmentation-Tracking Feedback:

Semantic masks gate SLAM feature selection, while high-precision camera poses warp segmentation masks forward temporally, reducing mask drift under object motion (Wang et al., 2018).

Bidirectional Feature Fusion:

In LiDAR–visual fusion, reconstructed 3D line directions from LiDAR are used as constraints in visual BA, while bundle-adjusted line directions are used to refine LiDAR ICP associations (Cao et al., 2023).

4. Quantitative Performance and Evaluation

Multiple bidirectionally coupled SLAM systems report significant performance gains over loosely coupled or unidirectional baselines, primarily in drift reduction, robustness, and geometric/photometric scene fidelity.

• DynaSLAM II: Tight coupling reduces mean translational RPE from 0.055 m/f (static-only) and 0.051 m/f (detect-and-mask) to 0.046 m/f (full coupling), with each direction of information contributing measurable accuracy gains (Bescos et al., 2020).
• GSO-SLAM: Achieves ATE RMSE 0.46 cm (Replica), 3.07 cm (TUM-RGBD), and 0.64 m (INS), with PSNR up to 34.48 dB and state-of-the-art geometric consistency, outperforming both naive coupled and decoupled methods in ablation studies (Yeon et al., 12 Feb 2026).
• iSLAM: Delivers relative drift reductions of about 22% (KITTI), a 40–60% improvement in ATE (EuRoC MAV), and consistently outperforms both classic and supervised deep visual-inertial SLAM pipelines, all without ground-truth supervision (Fu et al., 2023).
• LiDAR–visual: Coupled geometric feature fusion reduces mean translation error by 20–35% (vs. visual), 10–30% (vs. LiDAR). The system maintains operation under modality failure by bidirectional redundancy (Cao et al., 2023).

5. Practical Implementation and Data-Association Mechanisms

Bidirectional coupling mandates explicit, instance-level association and synchronization between data streams or modules.

• Instance segmentation and feature association: DynaSLAM II uses Mask-RCNN/YOLACT for per-frame object detection, labeling ORB features as static or dynamic for mapping and tracking (Bescos et al., 2020).
• Gaussian primitive initialization: GSO-SLAM replaces heuristics with a closed-form likelihood approach, directly initializing 2D Gaussian splats from sparse VO depth and image gradients, thereby accelerating convergence (Yeon et al., 12 Feb 2026).
• Sliding window optimization with recalibrated extrinsics: VIRAL SLAM propagates refined UWB-to-local transformations into the local state estimator thread, synchronizing all modalities (IMU, camera, lidar, UWB) to frame time (Nguyen et al., 2021).
• Feature matching and fusion: Visual–LiDAR systems match geometric line features by projecting both visual and LiDAR features into a spherical coordinate space, with efficient KD-tree search and direction averaging across modalities (Cao et al., 2023).
• Mask and map warping: Pose-driven warping of segmentation masks leverages refined SLAM pose estimates to compensate segmentation drift and increases label map precision (Wang et al., 2018).

6. Applications and Limitations

Bidirectionally coupled SLAM frameworks are deployed in autonomous driving, multi-robot collaboration, dense scene reconstruction, robotic navigation in visually degraded conditions, and real/augmented reality. They are particularly effective under:

• High dynamicity (moving agents, semantic scene changes),
• Multi-modal sensing (LiDAR, cameras, IMU, UWB),
• Dense or photorealistic geometric mapping.

Limitations include increased implementation complexity (scheduling, multithreading, marginalization), the need for reliable initialization and synchronization, potential for heuristic handover or switching logic between modalities, and computational cost of joint or alternating optimization (though several reported systems remain real-time even at high input rates) (Bescos et al., 2020, Nguyen et al., 2021, Wang et al., 2018, Cao et al., 2023).

Absence of full global loop closure in some frameworks allows unbounded drift over long traversals; manual extrinsic calibration or lack of semantic/intensity fusion are further acknowledged limitations (Cao et al., 2023). Bidirectional architectures offer natural points for further research such as joint probabilistic factor graphs, learned switching logic, and scalable, generalizable multi-agent settings.

7. Summary Table of Key Bidirectionally Coupled SLAM Frameworks

System	Coupled Modalities/Tasks	Optimization Approach
DynaSLAM II	Static SLAM & Multi-Object Tracking	Joint BA with dynamic/static
GSO-SLAM	Direct VO & Gaussian Scene Representation	Bidirectional EM (alt. BA)
iSLAM	Data-driven VO & Geometry-based Backend	Bilevel (meta) optimization
VIRAL SLAM	Camera, IMU, LiDAR, UWB	Local & global BA, feedback
LiDAR–Visual	Visual Lines & LiDAR 3D Features	Fusion + jointly updated BA
Seg.-Loc.	Instance Segmentation & Localization	Mutual gating and warping

Bidirectional coupling in SLAM represents a broad, powerful design principle, producing systems that robustly adapt to sensor, task, and scene changes by mutual constraint and information sharing between complementary subsystems (Bescos et al., 2020, Yeon et al., 12 Feb 2026, Fu et al., 2023, Nguyen et al., 2021, Cao et al., 2023, Wang et al., 2018).