⚡ electrical engineering

Belief-Adaptive Online Autonomy for Quadrotor UAV Navigation under GNSS Degradation in Urban Environments

This paper introduces a belief-adaptive online autonomy framework that enhances quadrotor navigation in GNSS-degraded urban environments by modeling GNSS trust as a latent state to proactively adjust sensor weighting and handle latency, thereby improving estimation accuracy and trajectory smoothness without requiring prior environmental knowledge.

Original authors: Deepak Kumar Panda, Weisi Guo

Published 2026-09-15

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Deepak Kumar Panda, Weisi Guo

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Flying a small drone through a dense city is a navigation problem of extreme difficulty. Unlike an open field where a drone can rely on signals from satellites orbiting overhead to know exactly where it is, a city canyon is a place of confusion. Tall buildings block the direct line of sight to the sky, and the hard surfaces of concrete and glass bounce those satellite signals around, creating echoes that arrive late and distorted. To a drone's computer, these reflections look like false positions, making the machine think it is somewhere it is not. When the drone tries to correct its course based on this bad information, it can drift off its intended path or, worse, crash into an obstacle it thought it had avoided. For these machines to operate safely in the future of urban air travel, they need a way to trust their sensors only when they are right, and to ignore them when the environment is lying to them.

Researchers at Cranfield University have developed a new method to help drones navigate these tricky conditions without needing to learn from years of past flight data. Their approach focuses on a specific type of intelligence: the ability to judge the reliability of a sensor in real time. In standard navigation systems, a drone fuses data from its internal motion sensors, which track movement very quickly, with data from the satellite system, which provides a global position but can be slow and error-prone in cities. Usually, the computer treats both sources as equally reliable, or adjusts its trust only after it has already made a mistake. The new framework changes this by giving the drone a "belief" about how much it should trust the satellite signal at any given moment. This belief is not a fixed setting; it is a living estimate that the drone updates continuously as it flies, allowing it to proactively ignore bad data before it causes a problem.

The core of this work is a system that treats the reliability of the satellite signal as a hidden state, much like a driver intuitively sensing that a road map is becoming unreliable as they enter a tunnel. The researchers created a mathematical model where this "trust" is represented by a single number that changes over time. When the drone detects that the satellite data is behaving strangely—perhaps because the signal is bouncing off a building or arriving late—the system does not just wait for the error to become large. Instead, it uses a sophisticated learning rule to adjust its confidence immediately. This rule looks at how the error is changing, not just the size of the error itself. By analyzing the shape of the mistake, the system can tell if the signal is likely to get worse or if it is just a temporary glitch, and it adjusts the drone's navigation accordingly. This allows the drone to rely more heavily on its internal motion sensors when the sky is blocked and switch back to the satellite signal as soon as the view clears, all without stopping to re-calculate its entire flight plan.

To test if this idea works, the team ran thousands of computer simulations of a drone flying through a virtual city filled with tall, rectangular obstacles. They programmed the simulation to mimic the real-world physics of signal blockage and the specific way satellite signals bounce off buildings, creating errors that change over time and depend on where the drone is located. They also introduced delays, simulating the lag that happens when a drone's computer takes a moment to process a signal. In these tests, they compared their new belief-adaptive system against older methods. The older methods, which simply react to errors after they happen, often led the drone to drift off course or follow a jagged, unstable path. The new system, however, kept the drone much closer to its intended route. The drone using the new method showed significantly less deviation from its planned path and maintained a smoother, more stable flight, even when the satellite signals were heavily corrupted by the urban environment.

The results of these simulations were clear and consistent. The new approach reduced the errors in the drone's estimated position by about eighty percent compared to the standard methods that do not adapt to changing conditions. It also cut the errors in tracking the flight path by thirty percent. This improvement came not from adding more sensors or making the drone heavier, but from a smarter way of processing the information it already had. The system managed to keep the drone's internal map of the world accurate by knowing exactly when to trust the satellite and when to doubt it. Crucially, this method does not require the drone to be trained on specific city layouts beforehand. It learns on the fly, making it suitable for any urban environment without the need for extensive pre-programming or offline training.

This work suggests that the key to safe urban flight lies in giving machines the ability to question their own data. By modeling the trustworthiness of a sensor as a dynamic belief that evolves with the environment, the researchers have created a navigation system that is both robust and efficient. The framework fits into existing flight control systems without requiring a complete overhaul of the hardware or software, meaning it could be integrated into current drone technology relatively easily. While the findings are currently based on computer simulations, they provide a strong foundation for future real-world tests. If these results hold up in actual flight trials, this method could be a vital step toward enabling the reliable, autonomous operation of drones in the complex and unpredictable airspace of our cities.

Technical Summary: Belief-Adaptive Online Autonomy for Quadrotor UAV Navigation under GNSS Degradation in Urban Environments

Problem Statement
Reliable online autonomy for quadrotor Unmanned Aerial Vehicles (UAVs) in Urban Air Mobility (UAM) is critically challenged by the degradation of Global Navigation Satellite Systems (GNSS) in dense urban environments. Unlike open-sky conditions, urban canyons introduce signal blockage, multipath reflections, interference, and latency. These factors generate non-stationary, temporally correlated, and geometry-dependent errors that violate the static noise assumptions of conventional GNSS–Inertial Measurement Unit (IMU) fusion filters (e.g., Extended Kalman Filters or EKF).

Existing adaptive filtering approaches often rely on reactive covariance tuning based on instantaneous innovation statistics. The authors argue that these methods fail to explicitly model measurement degradation, leading to delayed trust adaptation, inconsistent estimator confidence, and potential safety risks when large errors have already corrupted state estimates. Furthermore, machine-learning-based alternatives introduce training dependencies and computational overhead unsuitable for lightweight, data-efficient UAM operations.

Methodology
The paper proposes a belief-adaptive online autonomy framework that augments a standard EKF with an explicit, latent belief state representing GNSS reliability. The core components of the methodology include:

Latent Belief State Modeling:
- GNSS reliability is represented as a scalar belief parameter, $\beta_k$ , which modulates both the measurement covariance and the dynamics of the multipath bias process.
- This belief state is updated online without increasing the EKF state dimension, preserving real-time feasibility.
Curvature-Aware Second-Order Online Learning:
- Unlike first-order reactive methods, the framework employs a second-order online Newton update to adapt the belief state.
- A hybrid loss function, $\ell_k$ $ℓ_{k}$ , is minimized, combining:
  - Statistical Inconsistency: Penalized via the Normalized Innovation Squared (NIS) relative to a $\chi^2$ threshold.
  - State Stability: Penalized via the magnitude of fluctuations in the estimated position between time steps.
- The adaptation utilizes a dual-timescale curvature accumulation mechanism (fast and slow accumulators of gradient magnitudes) to approximate an online Newton step. This allows the system to be both reactive to abrupt changes and stable against oscillations.
- The belief parameter is mapped via a sigmoid function to ensure bounded values between 0 and 1.
Latency-Aware Out-of-Sequence Handling:
- The framework integrates a mechanism to handle delayed GNSS measurements (out-of-sequence updates).
- When delayed data arrives, the system performs a retroactive correction of the state estimate at the time of measurement and re-propagates the state forward to the current time, ensuring temporal consistency.
Degradation Simulation Model:
- The evaluation utilizes a controlled simulation environment where GNSS degradation is modeled as a function of the UAV's proximity to obstacles (Euclidean distance).
- Multipath bias is modeled as a first-order Gauss-Markov process with time constants and variance scaled by obstacle proximity.
- Latency is modeled as a stochastic, distance-dependent process.

Key Contributions
The paper explicitly claims the following contributions:

Curvature-Aware Second-Order Online Belief Adaptation: The introduction of a dual-timescale second-order update rule that explicitly models GNSS reliability as a latent belief state. This enables proactive trust modulation rather than reactive covariance tuning, improving dynamic navigation under non-stationary degradation.
Evaluation in Urban Air Mobility Environments: A comprehensive evaluation within a simulated UAM framework featuring geometry-dependent multipath bias and correlated latency. The study includes ablation studies comparing the proposed method against a standard EKF, a non-belief-aware adaptive EKF, and a first-order belief adaptation baseline.

Results
Simulations were conducted over 10 mission profiles in a 2D urban occupancy grid with a cascaded position-attitude controller and a global path planner. Key findings include:

Belief Convergence: The second-order learning approach achieved faster convergence to a stable trust level with significantly reduced oscillation compared to the first-order baseline, which exhibited persistent drift and higher variability.
Trajectory Tracking: In multipath-affected zones, the standard EKF deviated significantly from the planned path. The first-order method reduced deviation but showed delayed recovery. The proposed second-order method maintained closer adherence to the reference trajectory with smoother motion.
Error Reduction: Across 10 mission profiles, the proposed method demonstrated:
- An 80% reduction in estimation errors compared to the standard EKF.
- A 30% reduction in tracking errors compared to the standard EKF.
- Lower median errors and narrower error distributions (interquartile ranges) compared to standard, adaptive, and first-order methods, indicating improved consistency and robustness.

Significance and Claims
The paper positions this work as a solution that preserves the interpretability and structure of classical GNSS-IMU fusion pipelines while enabling robust online autonomy.

No Offline Training: The framework operates without offline training on environment-specific data, making it suitable for deployment in diverse urban settings.
Integration: It is designed to be integrated directly into existing flight-control pipelines without requiring new planners or controllers.
Proactive Adaptation: By explicitly modeling trust as a latent state updated via curvature-aware learning, the system avoids the "lag" inherent in reactive covariance tuning, providing stable sensor trust adjustment under geometry-dependent multipath and latency effects.

The authors conclude that while the current work focuses on the estimator adaptation layer within a controlled simulation, the framework lays the groundwork for future validation in real-world flight trials and multi-vehicle operations.

Technical Summary: Belief-Adaptive Online Autonomy for Quadrotor UAV Navigation under GNSS Degradation in Urban Environments

More like this