Answers are grounded in this paper and cite their sources.
In global navigation satellite system (GNSS)-denied environments, cooperative navigation among unmanned aerial vehicles (UAVs) under joint estimation relies on exchanging measurements and inertial measurement unit (IMU) data, because the joint estimator propagates the formation’s joint state and its cross-covariances. However, communication delays and out-of-order arrivals leave each vehicle’s estimate lagging behind its local IMU stream. To bring the estimate up to date, conventional re-mechanization replays buffered IMU frames one by one, which is computationally expensive. To avoid this frame-by-frame replay, we propose a factorized mechanization. Specifically, the Galilean-group decomposition isolates the navigation state at the interval start and confines IMU bias to a separate body-frame factor. As a result, the delayed estimate can be advanced in a single step per interval. Moreover, when the bias estimate changes, only this factor needs to be corrected using its Jacobian. To justify this correction, we derive a closed-form bound showing that its leading truncation error remains far below the IMU random-walk floor. A 600 s Monte Carlo stress test with inflated biases first shows that this simplification preserves accuracy: the final-epoch position RMSE relative to re-mechanization remains below 0.67 m, compared with 181 m for random-walk-only drift. Embedded benchmarks on a Cortex-M4F then quantify the computational gain: re-propagation is approximately 72× faster than re-mechanization; with an 8 s delay, total onboard processing is approximately 5.3× faster. In addition, each vehicle transmits 256 B per 1 s interval to the joint estimator instead of 4800 B. Finally, in a 695 s seven-UAV hybrid-formation test (three real and four simulated vehicles), the relative configuration error is 7.7 m, compared with 7.4 m for the no-delay baseline, 24 m for extended Kalman filter (EKF) linear propagation, and 29 m for state augmentation. By contrast, the shift methods diverge after approximately 386 s.
No takes yet. Share an insight, caveat, or question.
Liu et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: