GNSS/INS Ultra-Tight Coupling

• GNSS/INS ultra-tight coupling is an advanced fusion approach that directly integrates raw satellite signals and inertial data at deep levels to enhance navigation accuracy.
• It employs comprehensive state-space formulations, nonlinear measurement models, and optimization techniques to jointly estimate pose, velocity, clock biases, and integer ambiguities.
• The method significantly improves real-time robustness and achieves near-centimeter-level positioning in dynamic, interference-prone, and urban environments.

An ultra-tight coupling of Global Navigation Satellite System (GNSS) and inertial navigation system (INS) constitutes an advanced fusion architecture in which raw GNSS measurements—such as pseudorange, Doppler, and carrier-phase, as well as inertial sensor outputs—are jointly assimilated at deep integration levels in filtering or optimization frameworks. Unlike loose or classical tightly coupled schemes, ultra-tight strategies enable direct mutual assistance not only in navigation solutions but also at the receiver tracking loop and estimator model levels, thereby significantly improving robustness in interference-prone, highly dynamic, or urban environments.

1. State-Space Formulations in Ultra-Tight GNSS/INS Architectures

Ultra-tight GNSS/INS designs employ comprehensive state vectors unifying inertial and GNSS-specific parameters. For instance, in SRI-GVINS (Hu et al., 2024), the full state at epoch $k$ comprises:

$$x_k = [ x_S^\top,\, x_C^\top,\, x_P^\top,\, x_E^\top ]^\top$$

where:

• $x_S$ encodes inverse-depth parameters for tracked visual features,
• $x_C$ aggregates cloned IMU states (orientation quaternion $^Wq_{I_i}$ and position $^Wp_{I_i}$) over a sliding window,
• $x_P$ includes camera-to-IMU and GNSS-receiver-to-IMU extrinsics, multiconstellation clock biases $\delta t_j^r$, and clock drift $\delta\dot t^r$,
• $x_E$ comprises IMU velocity in the local frame, accelerometer bias $b_{a_k}$, and gyro bias $b_{g_k}$.

INS-relevant states reside in $x_C$ and $x_E$, while GNSS-specific states occupy bias entries in $x_P$, facilitating direct inference on navigation and signal-level parameters. This high-dimensional joint state enables full exploitation of measurement diversity (visual, inertial, multiple GNSS observables) and supports asynchronous sensor fusion by leveraging IMU pre-integration and marginalization strategies. Parallel frameworks, such as in GVINS (Cao et al., 2021) or optimization SLAM (Liu et al., 2020), align in jointly estimating pose, velocity, biases, clock offsets, and inter-frame transformations within sliding-window or graph-based optimizations.

2. Measurement Modeling and Nonlinear Fusion

Ultra-tight implementations rigorously formulate GNSS measurements using raw, undifferentiated observables as well as differenced carrier-phase and code measurements. In SRI-GVINS (Hu et al., 2024), distinct models are encoded for:

• GNSS pseudorange, incorporating geometric range, clock biases, and physical corrections (troposphere, ionosphere, Sagnac),
• Doppler shift, fusing LOS projections of satellite and receiver velocities, plus clock drift differentials,
• Single-differenced pseudorange and double-differenced carrier-phase, supporting RTK ambiguity resolution.

Each is linearized against current estimates, producing residual stacks combined into Jacobians $\mathbf{H}$ and noise covariances $\Sigma$. These are assimilated into the filter or optimizer (e.g., in a square-root inverse information filter or factor graph framework), inherently supporting measurement asynchronism via IMU pre-integration (Liu et al., 2020). Multi-sensor measurements (visual reprojection error, IMU factor residual, GNSS error terms) are minimized in joint cost functions, yielding estimates robust to signal interruptions and sensor dropouts. The same principle applies in tightly coupled RTK-INS MAP updates integrating double-differenced code/Doppler/phase with integer ambiguity states (Hu et al., 2024).

3. Filter and Optimization Algorithms for Ultra-Tight Coupling

The propagation and update steps leverage advanced filtering and nonlinear optimization strategies. The square-root inverse sliding-window filter (SRI-SWF) (Hu et al., 2024) maintains the information matrix in upper-triangular form $R$, avoiding explicit inversion and ensuring positive-definiteness under QR updates:

Propagation:

$R_{k+1|k} = \text{QR}(F \cdot R_{k|k})$

Measurement update:

$$\left[\begin{array}{cc} R_{k+1|k} & 0 \ \Sigma^{-1/2} H & \Sigma^{-1/2} \end{array}\right] \overset{\text{QR}}{\longrightarrow} \left[\begin{array}{cc} 0 & e \end{array}\right]$$

Variable initialization is handled via sequential filtering for reference-frame alignment (e.g., yaw and offset, marginalized once converged), RANSAC-based clock bias estimation, and online extrinsic calibration of GNSS-IMU lever arms using Doppler residual diversity (Hu et al., 2024). Nonlinear cost functions as in sliding-window factor graphs (Cao et al., 2021, Liu et al., 2020) and risk-averse MAP objectives (Hu et al., 2024) jointly optimize navigation, clock, and ambiguity states against measurement stacks.

Optimization-based schemes employ robust norms (Huber, LMI constraints), adaptive weighting (as in RAPS (Hu et al., 2024)), and marginalization for computational management.

4. RTK, Outlier Accommodation, and Integer Ambiguity Resolution

Ultra-tight fusion extends into real-time kinematic (RTK) carrier-phase processing, crucial for centimeter-level positioning. In risk-averse performance-specified architectures (Hu et al., 2024), double-differenced measurements and float ambiguities are integrated with a continuous measurement selection vector $b\in[0,1]^m$ to optimally down-weight outlier-prone satellites while guaranteeing required covariance bounds (SAE/ISO lane-level accuracy). If performance thresholds cannot be achieved (e.g., severe urban canyon), slack variables and penalties enforce graceful degradation.

Instantaneous RTK float solutions are produced each epoch (rather than relying on multi-epoch integer fixing and cycle-slip monitoring), with ambiguities treated as real-valued and subject to subsequent integer fixing if higher precision is required. Outlier detection and accommodation become an explicit part of the estimator, improving reliability in multipath-prone environments.

5. Integration at GNSS Receiver Baseband and Tracking Loop Level

A distinguishing feature of ultra-tight approaches is true integration at the GNSS baseband level. In vector-tracking loop architectures (Pages et al., 23 Jan 2026), inertial feedback directly modulates phase-locked loop (PLL) bandwidth and Doppler-rate commands per satellite channel:

• Adaptive loop filters receive Doppler-rate estimates from an INS-aided EKF, superseding classical frequency-locked loop assists.
• Observation generators package discriminator outputs into navigation observables for the estimator, which returns tracking commands to the receiver loops.
• Implementation on FPGA (e.g., Zynq UltraScale) realizes real-time updates and embedded resource efficiency, with negligible overhead relative to scalar tracking loops.

Similarly, smartphone-IMU-aided baseband SDRs (Luo et al., 2021) employ cascaded fusion pipelines: high-rate IMU drives strapdown mechanization, EKF fuses GNSS pseudorange/Doppler, and predicted observables directly assist code/carrier NCOs—achieving robust tracking continuity and carrier-phase stability even under low-cost consumer IMU input.

6. Computational and Numerical Characteristics

Ultra-tight filters (e.g., SRI-SWF (Hu et al., 2024)) achieve substantial computational savings through sparse QR decomposition and bounded-condition number operation, enabling full single-precision arithmetic and efficient embedded implementation. GNSS update steps—including high-complexity carrier-phase processing—are assimilated with minimal overhead relative to visual-inertial-only filtering (<3% extra CPU). Marginalization strategies surface in sliding-window optimizers (Liu et al., 2020), while vector-tracking FPGA deployments (Pages et al., 23 Jan 2026) maintain area budget parity with legacy scalar loop architectures.

7. Experimental Performance and Impact in Adverse Environments

Extensive evaluations demonstrate that ultra-tight GNSS/INS fusion yields quantifiable accuracy and robustness improvements: suppression of visual-inertial odometry drift, enhanced continuity in signal blockage, near-centimeter-level carrier-phase stability with even low-cost IMUs, and fulfillment (or explicit reporting) of lane-level horizontal and vertical error specifications over urban and canyon datasets (Hu et al., 2024, Hu et al., 2024, Luo et al., 2021). Performance metrics include horizontal error below 1.5 m (85.84% epochs), vertical error within 3 m (92.07% epochs) under smartphone-grade IMU in heavy urban environments (Hu et al., 2024), and comparative carrier-phase precision between consumer and high-grade IMUs (Luo et al., 2021).

Practical implications include applicability to autonomous ground vehicles, UAVs, and connected transportation systems requiring resilient, real-time geopositioning under GNSS outages, multipath, and high kinematic stress.

---

Ultra-tight GNSS/INS coupling, encompassing deep measurement modeling, joint state optimization, real-time tracking loop assistance, and robust outlier management, constitutes the current apex of sensor fusion for navigation integrity in GNSS-challenged scenarios. This paradigm demonstrates high measurement utilization, estimator resilience, and real-world computational efficiency substantiated across diverse and demanding empirical studies (Hu et al., 2024, Hu et al., 2024, Cao et al., 2021, Liu et al., 2020, Pages et al., 23 Jan 2026, Luo et al., 2021).