Solutions for GPS-Denied Navigation

Modern autonomous drones are often designed assuming GPS will always be available. In practice, however, many real-world missions take place in environments where satellite signals are weak, blocked, reflected, or intentionally jammed. Indoor facilities, dense urban areas, forests, tunnels, underground spaces, and disaster zones all create serious challenges for reliable positioning and navigation.

Why GPS-denied navigation matters

When GPS is unavailable, a drone can no longer rely on a stable global position estimate to follow routes, avoid obstacles, or complete its mission safely. In these situations, the system must estimate its motion, interpret the environment, and make flight decisions using onboard sensing and intelligent control.

A robust GPS-denied solution is especially important for applications such as:

• search and rescue in collapsed or smoke-filled areas

• warehouse and factory inspection

• forest exploration and environmental monitoring

• indoor logistics and inventory missions

• autonomous surveillance in cluttered environments

In all of these cases, navigation must remain stable, adaptive, and safe even when external positioning is lost.

A multi-agent architecture for autonomous navigation

One promising solution is to organize the drone intelligence as a federation of task-specific agents. Instead of relying on a single monolithic algorithm, the navigation system can be divided into specialized modules that cooperate through a shared internal state. This type of architecture improves modularity, interpretability, and robustness. Each agent focuses on one essential function, while the overall system integrates perception, localization, mapping, planning, and control.

1. Perception agent

The perception agent is responsible for understanding what the drone sees. Using onboard cameras and other sensors, it detects:

• obstacles

• structural features

• open corridors

• free space for safe flight

This agent provides the environmental awareness needed for navigation in cluttered scenes. In GPS-denied spaces, reliable perception becomes the first step toward safe autonomy.

2. Localization agent

Without GPS, the drone must estimate its own position using internal and visual measurements. The localization agent can combine information from:

• IMU data

• visual odometry

• lidar cues

• fiducial markers

• beacons or local anchors

A visual-inertial neural network is a strong option here, since it can fuse camera motion and inertial measurements to provide more accurate pose estimation under dynamic conditions.

3. Mapping agent

The mapping agent continuously updates a local representation of the environment. It identifies traversable regions, obstacle locations, and route constraints based on what the drone has observed so far. This local map does not need to be perfect or global. It only needs to be accurate enough to support short-horizon planning and safe navigation. In unfamiliar environments, this ability to build an online map is essential.

4. Planning agent

The planning agent decides how the drone should move. Based on the current mission goal and the shared environmental knowledge, it selects:

• path direction

• motion speed

• waypoint sequence

• fallback or recovery behaviors

This agent allows the drone to adapt its route when conditions change, obstacles appear, or uncertainty increases.

5. Safety and control agent

Even with good perception and planning, a drone still needs stable control. The safety and control agent ensures that the generated commands remain feasible and safe for the vehicle. It constrains aggressive actions, stabilizes the flight response, and handles disturbances. Advanced control strategies such as ADRC can improve robustness by compensating for model uncertainty and external perturbations, which is particularly valuable in real-world indoor or cluttered missions.

The role of a shared belief state

At the center of this architecture is a shared belief state. This internal representation acts as the common language between all agents. It can contain:

• estimated pose

• uncertainty level

• local map information

• mission intent

• confidence in observations

By updating this belief state continuously, the system enables all agents to work together in a coordinated way. Perception informs mapping, localization updates motion estimates, planning selects the next action, and control enforces safe execution.

Benefits of the proposed approach

A task-specific agent architecture offers several advantages for GPS-denied navigation:

Robustness: if one component becomes less reliable, the others can still support the mission.

Scalability: new sensing or reasoning modules can be added without redesigning the full system.

Adaptability: The drone can react to dynamic obstacles, uncertainty, or environmental changes in real time.

Safety: control constraints and stabilization strategies reduce the risk of collision or mission failure.

Interpretability: each module has a clear role, making the system easier to debug and improve.

Looking ahead

GPS-denied navigation is not just a backup mode. It is becoming a core capability for the next generation of autonomous drones. As missions move into more complex, dynamic, and constrained spaces, drones will need to rely increasingly on onboard intelligence rather than external infrastructure. The proposed solution, based on perception, localization, mapping, planning, and safety/control agents, provides a strong foundation for this future. By combining these components through a shared belief state, autonomous systems can achieve more reliable navigation even in environments where GPS is completely unavailable.

---

Efficient RF Signal Identification on Xilinx RFSoC

Edge radios are getting smarter. With Xilinx/Zynq UltraScale+ RFSoC, you can sample multi-GHz bandwidths, run DSP in programmable logic (PL), and execute deep nets on the ARM cores (PS) or an AI accelerator—all on a single board. This post lays out a practical, power-aware pipeline for real-time RF signal identification (modulation/class/protocol hints) that fits on RFSoC evaluation boards (e.g., ZCU111/ZCU208) or an RFSoC SOM. We’ll cover the end-to-end flow: streaming from RF ADCs → PL feature extraction → quantized inference → decisions and telemetry.

Why split this way? PL handles deterministic, high-rate DSP with microsecond latency. The PS/DPU runs compact, quantized models at tens of inferences per millisecond. DMA bridges them.

Physics-Informed Neural Network for Denoising Images Using Nonlinear PDE

Why PINNs for image denoising?

Classical denoisers—such as median filters, BM3D, and standard CNNs—do a decent job of smoothing noise but often blur edges or hallucinate textures. Physics-Informed Neural Networks (PINNs) add missing structure: they embed physical priors (from PDEs you’d actually use to describe diffusion or speckle statistics) directly into the training objective. The result is a model that not only learns patterns from data but also obeys a governing equation that favors realistic, edge-aware solutions.

Qualitative results of PDE-guided denoising. Left to right: Original, Noisy, and Denoised (PDE) images for three scenes (urban block, complex facility, aircraft on runway). The PDE-informed model suppresses speckle and additive noise while preserving edges and fine structures, recovering contrast and texture.

The core idea, in one picture

Think of denoising as finding a clean image u(x,y)u(x,y)u(x,y) that:

1. matches the observed image yyy after accounting for noise, and

2. satisfies a PDE prior that encodes how intensities should diffuse, sharpen, or stabilize.

Instead of only minimizing pixel loss ∥u−y∥\|u-y\|∥u−y∥, we add a PDE residual loss:

Which PDEs do we use?

Heat equation (linear diffusion)

A baseline smoother:

Great for reducing Gaussian noise, but risks oversmoothing edges.

MPMC (multi-phase / multi-component diffusion)

Useful when images contain distinct regions/phases (e.g., tissue types, materials). It regularizes piecewise-smooth areas and interfaces with tailored coupling terms.

Zhichang Guo (ZG) method for speckle

Speckle is multiplicative (common in SAR/ultrasound). The ZG family uses log-domain transforms and adaptive diffusion/regularization to suppress speckle while respecting radiometric statistics. In practice, we enforce a residual that stabilizes log-intensity variance and edge ratios consistent with speckle models.

Network backbones

We integrate the PDE losses into four families:

• UNet: strong skip connections, excellent for low-level restoration.

• ResUNet: residual blocks ease optimization for deeper models.

• U²-Net: “U-in-U” modules capture multi-scale structure with fewer parameters.

• Res2UNet: Res2 blocks split channels into granular groups, improving multi-scale feature interactions without big parameter growth.

  
  

Loss design (data + physics)

A typical training objective:

• Pixel/SSIM terms encourage fidelity.

• PDE residual promotes physics-consistent solutions.

• TV (optional) adds mild piecewise-smoothness.

Training details that matter

• Gradient stability: compute  via convolutional kernels to keep it GPU-friendly and stable.

• Boundary handling: reflective padding better matches physical imaging boundaries.

• Annealing: start with higher λpix\lambda_\text{pix}λpix​, softly increase λPDE\lambda_\text{PDE}λPDE​ as the network learns to reconstruct structure.

• Mixed precision: fine, but keep PDE ops in FP32 to avoid gradient underflow.

Limitations & future directions

• PDE choice matters: mismatched physics can bias results; validate per modality.

• Parameter sensitivity: κ\kappaκ, diffusion weights, and λPDE\lambda_{\text{PDE}}λPDE​ need tuning.

• Extensions: learned conductivity fields, plug-and-play priors, stochastic PDEs, and multi-task training (denoise + segmentation) are promising.

---

Takeaway

By marrying nonlinear PDE priors with deep networks, PINNs deliver denoisers that are both data-driven and physically grounded. In practice, that means higher PSNR/SSIM, better ENL/CNR for speckle, and—most importantly—clean images with sharp, trustworthy structure. If you’re already running a UNet-style pipeline, adding a PDE residual is a low-friction upgrade. For speckle-dominated imagery, fold in a ZG-style constraint and train in the log domain. You’ll get cleaner results without sacrificing the boundaries you care about.