MSC-LIO: Multi-State Constraint LiDAR-Inertial Odometry

• MSC-LIO is a tightly-coupled fusion framework that integrates LiDAR geometric constraints with high-rate IMU propagation for real-time ego-motion estimation.
• It employs a sliding-window optimization that leverages multi-frame point-to-plane residuals to robustly track and update pose estimates.
• State-of-the-art implementations demonstrate sub-decimeter to centimeter-level accuracy and extend the framework with degeneracy-aware strategies and auxiliary sensor fusion.

LiDAR-Inertial Odometry (MSC-LIO)

LiDAR-Inertial Odometry (LIO) refers to the simultaneous ego-motion estimation of a platform through the fusion of LiDAR and inertial measurement unit (IMU) data, exploiting the complementary properties of geometric environment constraints and high-rate inertial motion cues. Multi-State Constraint LiDAR-Inertial Odometry (MSC-LIO) designates a broad class of tightly-coupled fusion methods that employ a multi-state constraint Kalman filter (MSCKF) or fixed-lag smoothing strategy, maintaining a joint sliding window of recent poses and fusing multi-frame/inter-frame geometric constraints derived from LiDAR point clouds. These frameworks are characterized by principled treatment of high-rate IMU propagation, analytic point-to-plane (or higher-order) LiDAR measurement models, and a computational structure optimized for real-time operation on resource-limited platforms. State-of-the-art variants incorporate additional sensing modalities (e.g., UWB, camera, odometer), degeneracy-aware logic, and techniques for computational and statistical consistency (Zhang et al., 2024, Zhu et al., 13 Mar 2026, Zhang et al., 2024, Yuan et al., 2023).

1. Mathematical Foundations and State Representation

The canonical MSC-LIO filters maintain an augmented state vector, typically of dimension $15+6N$ (for $N$ keyframes), encoding the current IMU orientation $^G R_I \in SO(3)$, position $^G p_I \in \mathbb{R}^3$, velocity $v_I \in \mathbb{R}^3$, IMU biases ($b_g$, $b_a$), and $N$ “cloned” IMU or LiDAR pose states from recent time instants (Zhang et al., 2024, Zhang et al., 2024). The minimal error-state representation parameterizes small-angle SO(3) perturbations, while positions, velocities, and biases are realized as $\mathbb{R}^3$ increments. For systems with extrinsics (LiDAR–IMU, UWB–IMU), fixed or online-calibrated extrinsics are appended.

IMU propagation is governed by continuous-time kinematics: $$\dot{^G R_I} = ^G R_I [\omega_m - b_g - n_g]_\times,\quad \dot{v}_I = g^G + ^G R_I(a_m - b_a - n_a),\quad \dot{^G p_I} = v_I$$ with IMU bias drift modeled as random walks. Between updates, the error-state transitions are linearized, with covariance update: $N$0 where $N$1 arises from matrix exponentiation of the linearized dynamics over $N$2 (Zhang et al., 2024, Zhu et al., 13 Mar 2026). This framework enables efficient, filter-style real-time prediction.

2. LiDAR Measurement Models and Data Association

MSC-LIO implementations leverage direct geometric constraints from raw LiDAR point clouds, using multi-frame association for robustness and observability. The principal residual is the (possibly robustified) point-to-plane constraint: $N$3 where each LiDAR point is projected into a local coordinate system of a matching keyframe (via pose clones or map), and local neighborhoods are used to fit planes using PCA or cluster covariance analysis (Zhang et al., 2024, Zhu et al., 13 Mar 2026, Zhang et al., 2024).

Recent systems (e.g., (Zhu et al., 13 Mar 2026)) employ a lossless “cluster-to-plane” model: planar clusters in voxelized space serve as intermediate multi-state constraints. To avoid explicit parameterization or proliferation of feature states, null-space projection techniques are used to enforce “coplanarity” without overfitting, leading to consistent and low-bias estimation even in degenerate scenes.

For efficient online association, several approaches have been developed:

• Same-plane point tracking with minimal explicit feature extraction, coupling clusters over keyframes (Zhang et al., 2024)
• Multi-threaded, voxel-based data association (Zhu et al., 13 Mar 2026)
• Direct, feature-free distance field registration (Coto-Elena et al., 22 May 2025)

All produce analytic Jacobians with respect to the relevant state blocks, including IMU pose, extrinsics, and, if considered, LiDAR–IMU time delays (Zhang et al., 2024).

3. Sliding Window Fusion and Marginalization

MSC-LIO employs a sliding-window paradigm in which recent states (keyframes) are kept for direct multi-frame constraint construction. Upon processing a new LiDAR scan, the current pose is appended (“cloned”) to the state. Multivariate Kalman update or fixed-lag nonlinear optimization fuses stacked residuals from all active constraints: $N$4 After each update, old state clones are marginalized via Schur complement or block-matrix elimination, controlling both memory and computation (Zhang et al., 2024, Zhang et al., 2024, Zhu et al., 13 Mar 2026, Yuan et al., 2023). The modern trend is to exploit parallel implementation (e.g., plane fitting, voxel matching) to guarantee real-time feasibility on CPUs and edge devices.

Logical or “semi-elastic” constraints are introduced in (Yuan et al., 2023) to improve temporal consistency—allowing slight relaxation between sequential states to avoid inconsistency from model misalignment, balancing rigidity and flexibility in the optimization.

4. Extensions: Degeneracy Handling, Auxiliary Modalities, and Robustness

Environments with poor LiDAR constraint observability (corridors, single-plane, or NLOS) challenge the consistency of LIO. Degeneracy-aware approaches incorporate explicit monitoring (e.g., Hessian eigenvalue analysis) to detect under-constrained directions (Han et al., 2023, Kim et al., 3 Apr 2026). When detected, auxiliary sensing (wheel odometry, vision, UWB, or learned priors) can be injected as additional constraints only as needed, provably lowering the covariance bound according to the Cramer-Rao lower bound (Han et al., 2023).

Hybrid frameworks integrate, e.g., ultra-wideband ranging (as in MR-ULINS (Zhang et al., 2024)), deep neural network velocity priors (e.g., ALIVE-LIO (Kim et al., 3 Apr 2026)), or explicit SE(2)–motion constraints for ground vehicles (SE2LIO (Chen et al., 2023)). All are realized within the MSC-LIO architecture, only augmenting the state and residuals correspondingly. Sensors are tightly time-synchronized and spatially calibrated (sometimes with adaptive re-calibration included) (Qingqing et al., 2023).

Robust data-association, e.g., via multi-epoch outlier rejection or robustified loss/covariance modeling (Huber, Cauchy, adaptive weighting), further improves operation in the presence of outliers or sensor artifacts (Zhang et al., 2024, Coto-Elena et al., 22 May 2025, Malladi et al., 8 Sep 2025).

5. Algorithmic Pipelines and Runtime Analysis

A generic MSC-LIO pipeline proceeds as follows: IMU-driven state propagation at high rate; LiDAR scan arrival triggers deskewing/motion correction; geometric constraints are constructed and associated (planes, clusters, or distance fields); multi-state residuals are assembled and stacked; joint filter (Kalman, ESKF) or optimization (Gauss–Newton) update is performed; new state clones are introduced; marginalized priors keep the problem tractable (Zhang et al., 2024, Yuan et al., 2023, Coto-Elena et al., 22 May 2025).

Experiments consistently demonstrate <0.1 m RMS ATE in structured scenes, sub-decimeter to centimeter-level drift across long indoor and outdoor runs, and robust operation in degenerate or NLOS environments. Ablation studies indicate superior accuracy and efficiency compared to pure frame-to-frame or naive graph-based approaches, with edge-device implementations demonstrating 20–100 ms per scan update (well within LiDAR scan periods) (Zhang et al., 2024, Zhang et al., 2024, Zhu et al., 13 Mar 2026).

6. Current Benchmarks, Limitations, and Design Insights

Recent MSC-LIO systems outperform both traditional LIO and map-based least-squares systems across public and private datasets, with demonstrable gains in efficiency (parallelizable data-association and update), robustness (degeneracy-aware fusion), and theoretical consistency (low NEES, avoidance of overconfidence) (Zhu et al., 13 Mar 2026, Zhang et al., 2024). Tabled comparisons consistently show state-of-the-art error metrics and runtime on both desktop and embedded platforms.

However, limitations persist:

• Parametric constraints (e.g., SE(2)) may suffer under rapidly-varying ground slopes or unmodeled 3D vehicle motion (Chen et al., 2023).
• Robustness depends on reliable extrinsic/time delay calibration and dynamic adaptation in challenging real-world deployment (Qingqing et al., 2023, Kim et al., 3 Apr 2026).
• Map scale management and computational scalability remain bottlenecks for extremely large or high-density environments (Coto-Elena et al., 22 May 2025).

Open research directions include more sophisticated degeneracy detection and adaptive fusion heuristics, integration of alternative sensor modalities (cameras, radar), dynamic ground-plane estimation for non-planar navigation, and joint end-to-end optimization frameworks for fully consistent multi-sensor state estimation.

• (Zhang et al., 2024) MSC-LIO: An MSCKF-Based LiDAR-Inertial Odometry with Same-Plane Cluster Tracking
• (Zhang et al., 2024) MR-ULINS: A Tightly-Coupled UWB-LiDAR-Inertial Estimator with Multi-Epoch Outlier Rejection
• (Zhu et al., 13 Mar 2026) Consistent and Efficient MSCKF-based LiDAR-Inertial Odometry with Inferred Cluster-to-Plane Constraints for UAVs
• (Yuan et al., 2023) Semi-Elastic LiDAR-Inertial Odometry
• (Han et al., 2023) DAMS-LIO: A Degeneration-Aware and Modular Sensor-Fusion LiDAR-inertial Odometry
• (Kim et al., 3 Apr 2026) ALIVE-LIO: Degeneracy-Aware Learning of Inertial Velocity for Enhancing ESKF-Based LiDAR-Inertial Odometry
• (Malladi et al., 8 Sep 2025) A Robust Approach for LiDAR-Inertial Odometry Without Sensor-Specific Modeling
• (Coto-Elena et al., 22 May 2025) D-LIO: 6DoF Direct LiDAR-Inertial Odometry based on Simultaneous Truncated Distance Field Mapping
• (Chen et al., 2023) Versatile LiDAR-Inertial Odometry With SE (2) Constraints for Ground Vehicles
• (Qingqing et al., 2023) Robust Multi-Modal Multi-LiDAR-Inertial Odometry and Mapping for Indoor Environments
• (Ye et al., 2019) Tightly Coupled 3D Lidar Inertial Odometry and Mapping