Fusion Architectures: Loose, Tight, and Deep Coupling Explained

Introduction

Choosing how to integrate GNSS, IMU, and other sensors is one of the most consequential design decisions in a navigation system. The architecture determines not only what performance is achievable but also what hardware access is required, what computational resources are needed, and how the system behaves in degraded conditions. Three principal fusion architectures - loosely coupled, tightly coupled, and deeply coupled - represent a progression from simple to complex, with corresponding improvements in performance and robustness.

Loosely Coupled Architecture

In loosely coupled (LC) integration, the GNSS receiver and the INS each operate as independent subsystems with their own complete navigation solutions. The GNSS receiver outputs its standard position and velocity fix, and these are treated as external measurements by a fusion Kalman filter running alongside the INS.

The fusion filter estimates the errors in the INS solution by comparing the INS-derived position and velocity against the GNSS-derived position and velocity. When a GNSS fix is available, the filter generates correction terms; when GNSS is unavailable, the filter propagates on inertial data alone.

Advantages of Loosely Coupled

• Works with any GNSS receiver that outputs standard NMEA or binary position/velocity - no access to raw measurements is needed
• Simple to implement - the GNSS receiver is a black box
• Mature, well-understood algorithms widely available in open-source toolkits
• Straightforward to debug and validate

Limitations of Loosely Coupled

• Requires at least four satellites for the GNSS receiver to generate a valid position fix - below four satellites, no GNSS update is available
• The GNSS receiver''s internal filtering may smooth out real signal anomalies, hiding problems from the fusion filter
• Correlated errors between the GNSS receiver and the fusion filter can degrade filter performance in some configurations

Tightly Coupled Architecture

Tightly coupled (TC) integration bypasses the GNSS receiver''s position computation entirely. Instead, raw GNSS measurements - pseudoranges and pseudorange rates from each tracked satellite - are fed directly into the fusion Kalman filter alongside the INS state. The filter simultaneously estimates INS errors, receiver clock bias, and clock drift rate, using the pseudorange residuals (measured minus predicted) as the correction mechanism.

Predicted pseudoranges are computed from the current INS position estimate and the satellite positions derived from broadcast ephemeris. The difference between measured and predicted pseudoranges drives the filter update. Because each satellite contributes an independent measurement regardless of the total count, TC integration can maintain GNSS-assisted navigation with as few as one or two visible satellites.

Advantages of Tightly Coupled

• Effective with fewer than four satellites - partial sky visibility still provides valuable constraints
• The fusion filter can detect and reject bad satellite measurements directly
• Better performance in urban canyons and satellite-sparse environments
• INS can assist GNSS signal re-acquisition after short outages

Limitations of Tightly Coupled

• Requires access to raw pseudorange measurements - not all receivers expose this data
• Significantly more complex to implement than loose coupling
• Requires satellite ephemeris data to compute predicted ranges
• More complex filter with larger state vector (17 states vs 15 for loosely coupled)

Deeply Coupled (Ultra-Tight) Architecture

Deep integration, also called ultra-tight coupling, extends the fusion to the signal processing level within the GNSS receiver. Rather than extracting pseudoranges and feeding them to an external filter, the INS state is used directly to assist the receiver''s carrier tracking loops and code tracking loops. The INS-derived velocity and line-of-sight geometry enables the receiver to predict the expected Doppler shift for each satellite, tightening the tracking loop bandwidth and enabling signal tracking in conditions - high dynamics, near-jamming - that would otherwise cause loss of lock.

This architecture delivers the highest performance but requires deep integration between the GNSS receiver hardware/firmware and the INS processing. It is primarily found in military systems, precision weapons, and specialised high-performance commercial navigation systems.

Centralised vs Federated Kalman Filters

When multiple sensor modalities beyond GNSS and IMU are included - such as LiDAR, cameras, or radar - the system architecture must decide how to structure the fusion. A centralised Kalman filter fuses all measurements from all sensors into a single large filter. This is theoretically optimal but computationally intensive and fragile - a single sensor failure can corrupt the entire state estimate.

A federated architecture uses separate local filters for each sensor pair (e.g., GNSS/IMU, LiDAR/IMU), with a master filter that combines the local estimates. Federated architectures are more fault-tolerant and computationally tractable, at the cost of some optimality. They are the preferred architecture for safety-critical systems where sensor fault isolation is essential.

Practical Implementation Considerations

Regardless of architecture, successful GNSS/INS fusion requires careful attention to time synchronisation - GNSS and IMU measurements must be precisely timestamped to a common reference to avoid introducing artificial errors from temporal misalignment. Lever arm compensation - accounting for the physical offset between the GNSS antenna phase centre and the IMU sensing centre - is another critical calibration step that is often overlooked in practice but can introduce systematic errors of tens of centimetres if neglected.