Bavovna AI
Project highlights
About the client
About the client
Bavovna AI is a tech startup developing a dual-use, AI-driven alternative navigation system for Uncrewed Aerial Vehicles (UAVs) operating in GNSS-denied (Global Navigation Satellite System) and electronic warfare-threatened environments.
In modern warfare, GNSS signals like GPS are often jammed by enemy forces, leaving UAVs vulnerable, unable to navigate or complete their missions. This problem is especially critical for Ukraine, now in its fourth year of defending its sovereignty against russian aggression.
To support Ukraine’s fight for independence, Bavovna AI launched a project to build a reliable navigation system that keeps drones on mission, even when satellite signals are lost. At the same time, the technology holds strong potential beyond the battlefield. From tunnels to infrastructure inspections in remote or GPS-blocked areas, Bavovna’s solution opens doors to safer, smarter UAV operations across a range of commercial sectors.
CLIENT REQUEST
Bavovna AI partnered with Flyaps to build an accurate navigation system for UAVs operating in GNSS-denied environments.
Bavovna AI came to us with a clear vision for their unmanned drone navigation system. They had the theory and detailed documentation on how to implement it, but lacked the team to build the actual system.
At the heart of their concept was a custom AI model trained on the telemetric data, which includes real readings from drone sensors during flight. Bavovna worked with partners (Aurelia, Nemesis), who could supply this data, but it was delivered manually. After each mission, drone operators had to transfer flight data via cables to a computer, upload it to Google Drive, and then email the links to the team working on the AI model. The process was slow, posing a serious risk to the project timeline.
Another challenging step in AI model development was testing. We had to test how our AI model processed telemetric data to make navigation decisions in real time. Without a controlled environment to validate predictions before sending drones into flight, the risk of failure remained high. We needed a way to test and refine the system before taking it live, because when you’re navigating without GPS, there’s no room for guesswork.
problem
To build automated UAV navigation for GNSS-denied environments, we had to solve several challenges:
Before GPS became standard, aerial vehicles used Inertial Navigation Systems (INS) to track their movement. These systems rely on ultra-precise sensors and complex algorithms to calculate position based on acceleration and rotation. While effective, INS is too expensive, especially in frontline drone operations in Ukraine, where hostilities are estimated to be over 1,200 kilometers long.
Like in many other industries, AI offered a promising alternative. AI could enhance INS performance, reduce the need for costly calibration, and high-end aerospace hardware. With an AI model, specifically a Recurrent Neural Network, and raw sensor data, like readings from barometer, accelerometer, and gyroscope, a UAV could accurately navigate even in GNSS-denied environments
At Flyaps, we’ve developed a range of AI-powered solutions, from AI-powered marketplaces to UrbanSim, an ML-ready platform for urban planning. With that expertise, we confidently took on Bavovna’s project to build a GPS-free navigation system for drones.
approach
To enable precise unmanned navigation, we developed an RNN model that uses raw sensor data to predict a UAV’s position without GNSS coverage.
solution
To achieve high accuracy and reliable GPS-free navigation, we built a neural network model based on theoretical documentation proven effective in real-world scenarios. The entire system runs on a cloud-based infrastructure using a self-hosted Kubernetes cluster. This setup enables in-cluster processing of all telemetry data. Once a UAV completes its flight and connects to the Internet, it automatically uploads its telemetry data to the cloud. From there, the data is stored and structured to improve the model accuracy in the future.
Here is how GPS-free navigation works in simple terms.
The whole project took about seven months and started with gathering training data.
To train the model, we required around 100 hours of flight recordings. Bavovna AI connected us with companies that supplied the telemetry. We designed specific flight scenarios to gather the training data and analyzed it to determine its suitability for training our AI model. In parallel, we were studying research papers provided by the client.
At first, data collection was a mess, slow, manual, and far from scalable. After every mission, drone operators had to transfer flight data via cables to a computer, upload it to Google Drive, and then email us the links. It was a tedious chain of steps, and with hours of flight time piling up, the amount of raw data quickly became overwhelming.
To fix that, we automated the entire data pipeline. As soon as a flight ended, the data was automatically uploaded to the cloud, formatted correctly, and made ready for AI model training.
We designed an MLOps pipeline to automate the entire model lifecycle. As new telemetry data was uploaded from drone flights, the system would automatically process and add it to the training dataset. This would allow us to regularly retrain the AI model with fresh data, keeping its accuracy and stability high. The concept also included continuous monitoring of the model’s performance to spot any changes in flight conditions or sensor behavior, triggering retraining whenever necessary.
Before testing the UAV in real-world scenarios, we used an out-of-the-box virtual simulator to assess how the navigation model performed. When the model demonstrated accurate results in the simulator, we proceeded to field testing at an actual test site in Poland. There, we deployed real drones equipped with the full hardware setup to validate the system under real flight conditions.
During these tests, the drone successfully flew autonomously, collected sensor data throughout the flight, and, upon returning to base, automatically uploaded the recorded telemetry to our cloud-based data ingestion system for further analysis and model improvement.
Result
Currently based in the US, Bavovna AI has raised $2.7 million in seed-stage funding. In January 2025, the company successfully demonstrated its Hybrid Inertial Navigation System (H-INS) to U.S. military experts in Tampa, FL. The system autonomously completed a 7.8 km mission without GPS, maps, or visual odometry, landing precisely with 99.98% accuracy.
Need AI that transforms your industry? We’re here to help. We provide data engineering, AI development, and MLOps services.
Let’s talk
