A peer-reviewed study published in Satellite Navigation describes an AI framework that can estimate a drone’s position, including latitude, longitude, and elevation, without relying on GPS signals or any camera input. The system, called CLAK, instead fuses data from LiDAR, barometric altitude sensors, and inertial measurement units to locate an unmanned aerial vehicle in environments where satellite signals are jammed, spoofed, or simply unavailable. The research arrives as militaries, disaster-response agencies, and commercial operators increasingly confront the limits of GPS-dependent flight in contested or indoor settings.
What is verified so far
CLAK stands for CNN-LSTM-Attention and Kolmogorov-Arnold Networks, a layered machine-learning architecture that processes non-visual sensor streams to produce a geographic fix. According to the Satellite Navigation report, the system builds a 17-dimensional fused feature vector by aligning data from sensors that operate at different sampling rates. LiDAR returns, barometric pressure readings, and IMU acceleration and rotation measurements each arrive at their own cadence. CLAK synchronizes these streams before feeding them through convolutional neural networks for spatial pattern extraction, long short-term memory layers for temporal sequence learning, attention mechanisms for weighting the most informative features, and Kolmogorov-Arnold Networks for final regression to geographic coordinates.
The evaluation so far is entirely simulation-based. Researchers tested CLAK inside a synthetic environment built on ROS2, PX4, Gazebo, and QGC, using the Taif digital elevation model as terrain ground truth, as described in a release on the simulation setup. That stack is standard in drone research: ROS2 handles middleware communication, PX4 runs the flight controller, Gazebo renders the physics world, and QGC provides a ground-control interface. The choice of the Taif DEM, which represents a mountainous region in Saudi Arabia, suggests the team wanted terrain with enough elevation variation to stress-test barometric and LiDAR-based altitude estimates.
No real-world flight data has been published for CLAK. Every accuracy figure reported in the paper comes from these synthetic runs. That distinction matters because simulation environments offer clean sensor noise profiles, perfect timing, and no wind gusts, radio interference, or structural occlusions. Translating simulation-grade performance to an actual airframe in a GPS-denied urban canyon or underground mine is a separate engineering challenge that the authors have not yet demonstrated.
The peer-reviewed version of the work is also accessible through a DOI-backed publication, which confirms the core claims about using LiDAR, barometric, and inertial data to replace satellite navigation under jamming or signal loss. Together, these sources establish that CLAK is a documented academic contribution rather than a marketing concept or unreviewed preprint.
Parallel efforts strip away other sensors too
CLAK is not the only recent project rethinking which sensors a drone truly needs. A separate study in npj Robotics took the opposite approach: instead of discarding cameras and keeping inertial sensors, that team discarded the IMU and kept only an event camera for vision-based attitude estimation and control. In that work, asynchronous brightness changes from the event camera feed a learning-based controller that stabilizes and steers the aircraft without conventional gyroscopes or accelerometers.
The same vision-only research can also be reached via a publisher access gateway, underscoring that it, too, has passed formal peer review. The contrast with CLAK is instructive. One group bets that non-visual signals alone can replace satellite positioning; the other bets that vision alone can replace inertial sensing. Both are testing how far a single sensor modality, enhanced by AI, can stretch before accuracy degrades beyond usefulness.
A third line of work, MiFly from MIT and Cornell, tackles the same GPS-denied problem with millimeter-wave radio frequency signals. MiFly achieves six-degree-of-freedom self-localization using a single mmWave anchor, making it viable in dark, smoke-filled, or visually occluded indoor spaces where neither cameras nor LiDAR perform well. The method relies on phase and amplitude information from reflected radio waves to infer both position and orientation relative to the anchor.
Taken together, these three projects reveal a broader pattern: research teams are systematically eliminating single points of failure in drone navigation. GPS denial is the most obvious threat, but camera failure from dust, darkness, or damage is equally disabling. Each project removes a different sensor dependency and replaces it with AI-driven inference from whatever signals remain available, whether that is terrain-matching LiDAR, sparse visual events, or radio reflections.
What remains uncertain
The biggest open question for CLAK is whether its simulation results will hold up on physical hardware. Synthetic ROS2/PX4 environments can model sensor noise, but they rarely capture the full complexity of real-world interference. LiDAR returns in a dusty post-earthquake environment, for example, scatter differently than in a clean Gazebo render. Barometric altitude readings drift with weather fronts and temperature gradients. IMU data accumulates integration error over time, especially under vibration.
The Satellite Navigation paper does not describe any plan or timeline for flight tests on an actual UAV platform, and there is no public dataset showing how CLAK responds to sensor dropouts, partial occlusions, or dynamic obstacles. Without such experiments, it is difficult to know whether the attention mechanism and Kolmogorov-Arnold Networks can gracefully down-weight corrupted signals or whether they will propagate errors into the final position estimate.
There is also no unified benchmark comparing CLAK, MiFly, and the vision-only attitude system against each other. Each team tested under different conditions, with different metrics, and against different baselines. CLAK targets outdoor GPS-denied localization using terrain-matching signals tied to a digital elevation model. MiFly targets indoor positioning with a single radio anchor and emphasizes robustness in visually degraded conditions. The npj Robotics work targets attitude control, not geographic localization at all, focusing on how a drone maintains orientation rather than where it is on a map.
Readers should therefore resist the temptation to rank these systems by “accuracy” based on headline numbers alone. No published study has yet tested them head-to-head or in combination, and there is no shared benchmark suite that would allow a fair comparison across outdoor, indoor, and vision-only regimes. Any claims that one approach is definitively superior would go beyond the available evidence.
Direct researcher commentary is also thin. The available institutional releases, including a summary aimed at general audiences, provide simplified descriptions of the CLAK methodology but do not include extended quotations from the lead authors explaining design tradeoffs, failure modes, or next steps. Without those statements, outside observers are left interpreting the paper’s technical sections without the researchers’ own framing of what they consider the system’s weakest link or most promising application.
How to read the evidence
The strongest evidence here is the peer-reviewed CLAK paper, which documents the architecture, training procedure, and simulation-based evaluation in detail. That record supports the claim that AI can fuse LiDAR, barometric, and inertial data to estimate a drone’s position without GPS or cameras in a controlled virtual environment. The supporting simulation description and institutional summaries corroborate the high-level narrative but add little in the way of independent validation.
By contrast, the event-camera attitude controller and the MiFly mmWave system demonstrate that similar AI-heavy approaches can work on physical hardware, at least in constrained tests. However, they address different aspects of navigation and operate under different assumptions about the environment and available infrastructure. Drawing broad conclusions about “AI replacing GPS” from any single project would be premature.
A cautious reading treats CLAK as an early proof-of-concept for GPS-free, non-visual localization that still needs rigorous field trials. The most justified inferences are that multi-sensor fusion with deep learning can exploit subtle correlations in non-visual data, and that simulation tools now make it relatively straightforward to prototype such systems. Claims about battlefield readiness, disaster-response deployment, or commercial certification would require evidence that does not yet exist in the public record.
For now, the safest takeaway is that drone navigation research is rapidly diversifying away from a single point of failure. Whether future aircraft rely on terrain-scanning LiDAR, event-based vision, radio beacons, or some hybrid of all three, the common thread is an attempt to keep flying when GPS and conventional cameras cannot be trusted. CLAK is a notable entry in that trend, but its real test will come when its algorithms leave the simulator and confront the messy physics of the real world.
More from Morning Overview
*This article was researched with the help of AI, with human editors creating the final content.
