Autonomous Drone Systems for Critical Operations

Using the latest research on computer vision (June 2026), we have developed a novel neural network system, capable of matching the accuracy of much larger state of the art models such as the Yolov11 and Yolov26 in various drone target detection and tiny person detection tasks. More importantly, our model could operate purely on our CPU processing stack and still have much higher frames per second (40+FPS) than the larger benchmark models which were already running on top of the line edge GPU hardware such as the current market best, the Hailo AI Accelerator module series. So not only can we achieve the same level of performance in accuracy and even higher FPS, but we do so with less supply chain risks and hardware costs since we do require a GPU. We also have a 2nd detection neural network model of the same architecture but this time for the thermal vision camera. This grants us flexibility in a fire and forget scenario. These two together we call the IXEAN MATA system, which means Eye.  

MATA architecture neural networks can range from extremely lightweight models for the most resource constrained of tasks to heavier but still efficient models that prioritize high accuracy. The lightweight models are suitable for smaller robotic platforms that have even more limited compute capacity or on drones that require extremely high FPS for tracking. Similarly the larger models are suitable for demanding computer vision tasks such as multiple target detection, classification, and tracking.

To operate in GPS denied environments and to counter other forms of electronic warfare, we also house locally on board a 3rd neural network with a custom extended kalman filter.  This is our novel neural inertial odometry system that allows the drone to estimate its GPS position at all times with high accuracy using just inertial sensor measurements such as those from the accelerometer. We call this the IXEAN LANDAS which stands for Pathway.  The LANDAS builds off on and addresses the many issues of the latest inertial odometry literature.  What sets our system apart from commonly used state of the art odometric systems is that these other systems often require heavy use of either LIDAR  or vision based sensors (visual inertial odometry). We achieve at par local and global position estimation with these benchmark visual inertial odometry systems with much less hardware use. With our LANDAS system, we can return to home fully autonomously even if the GPS module is completely jammed. This also allows us to fly to predefined way points that are potentially dense with electronic warfare countermeasures. High accuracy pure inertial odometry systems are rare and extremely lightweight and low cost.

In the plots above, we can see that the estimated velocities and trajectories of IXEAN LANDAS versus the actual ground truth. It is capable of closely following the movements of the drone even for extended periods of time.

Our autonomous navigation, tracking software, swarm intelligence, and graphical user interface (GUI) software combined we refer to as IXEAN DIWA, which stands for Mind. Utilizing compressed and quantized data from MATA and LANDAS, our autonomous navigation system is lightweight and runs on a C++ ROS2 framework.  DIWA takes a variety of forms, albeit all unified by the same core operating system. From target interception algorithms as in Keris to search, patrol, and track algorithms as in Lawin. More importantly, we utilize a unique swarm intelligence algorithm that allows multiple drone units to coordinate with each other's autonomous navigation systems allowing for more accurate pathfinding and target tracking. DIWA Swarm varies in its form depending on the robotic platform. It can range from very simple notifications to other drones in the network whether it has a detection or not such as in the UWAK drone, to a complete swarm intelligence guided autonomous navigation such as in LAWIN where each drone can summon other drones to follow and assist in searching a particular region. We also have a custom general user interface that allows operators ease of use in customizing DIWA and controlling it throughout the entirety of the operation. We build off the latest targeting and tracking algorithms, implementing a number of novel extensions on them such that even when the target is not on screen, we can estimate its most probable location and continue tracking it, ensuring persistent interception or tracking.