Aircraft autopilot systems have been around since the early 20th century. Modern autopilot systems effectively support pilots at every phase of the flight, enabling greater safety, efficiency, and comfort. Unmanned aerial vehicles (UAVs) have even more advanced navigation and safety systems. 

But how technologically far are we from achieving full autonomy? Given the latest advances in sensor fusion, computer vision, deep and reinforcement learning, reliable autonomous navigation systems will become a reality once we resolve the last cluster of challenges.  

5 Challenges of Creating Autonomous Navigation Systems for UAVs 

Autonomous UAVs have many great use cases — from urban package delivery to automated industrial asset inspections, border patrols, or land surveying missions. What prevents the technology from going mainstream is the reliability requirements. Autonomous navigation systems must be ultra-safe with minimal possibility of error. To make that happen, industry leaders are working on solving the next five problems. 

Dynamic Environmental Awareness 

UAVs need continuous spatial awareness about flight conditions and obstacles to operate autonomously without posing risks to people and infrastructure. Modern drones already include advanced sensing capabilities, ranging from HD cameras to LiDAR and optical flow sensors. Computer vision and navigation algorithms can process all of these data points on-device to provide real-time environmental awareness to drones for autonomous operations. 

The Fly4Future team recently presented INEEGO — an indoor inspection drone with an autonomous navigation system. The drone can effectively glide through the premises using only a partially known map and adjust its path based on the data from onboard sensors to avoid obstacles. With INEEGO, pilots can seamlessly inspect AC pipelines, carrier beams, and other industrial structures for signs of deterioration with little risk of collision. 

GPS Denied Environments

GPS has long been a bottleneck in UAV reliability. If the signal is down or unreliable (which happens often in dense urban areas), the drone loses all positioning navigation. Drone signals can also be jammed or spoofed, further affecting the safety of autonomous flights. 

Soundly, there are alternatives to GNSS connectivity. US manufacturer Bavovna developed a hybrid AI navigation system for operating in GPS-denied/compromised environments. The company’s flagship product is a SWAP,  low-cost, modular solution, combining an onboard hardware processing unit and pre-trained AI algorithms to provide precise Position, Navigation, and Timing (PNT), crucial for autonomous operations. With Bavovna’s aerial autonomous navigation systems, UAVs can fly fully autonomous complex missions, relying exclusively on onboard sensors like IMU array (accelerometer, gyroscope), airflow sensor, compass, and barometer, among others. 

Battery Management 

Drones have limited battery capacity. If an autonomous UAV runs out of charge mid-route and fails to safely land, that’s no good. So, researchers are evaluating different options for improving drone battery management. 

NTIS Research Centre, for example, created an experimental mechatronic system for automatic drone battery management. Droneport can autonomously swap batteries on UAVs without any human intervention.  Compact and easily assembled, the Droneport robo-arm can perform scheduled battery swapping tasks on drones, equipped with a specialized battery holding case with high accuracy. 

Self-charging drones are another actively explored option. Drones4Safety technology, developed in Denmark, allows charging UAVs using railway and power line cables. When a drone reaches a low battery status, an autonomous navigation systems navigates it toward the nearest overhead line for charging, using data from GNSS and EGNOS. On-board sensors also provide extra detection and navigation towards the nearest power lines. Several field trials have already proven the viability of such an approach. However, the team still needs to work on the systems’ accuracy to avoid potential damage to the power lines. 

Autonomous Take-Off and Landing

Present-day drones require a pilot to facilitate take-off and landing. In autonomous operations, drones will have to auto-locate suitable landing sites and ensure that no obstacles or humans are in the way of a safe takeoff/landing. Again, this problem requires improvements to current navigation capabilities and some upgrades to hardware design. 

Evolve Dynamics is working in this direction. The manufacturer recently showcased its 

Sky Mantis UAV can perform fully autonomous landing, loitering, and zonal position holds with high accuracy. The model relies on ground-based radar beacons communicating with a Sensoriis airside radar mounted on the Sky Mantis UAV, which is used to make decisions based on precision positioning data.

A group of Polish researchers, in turn, proposed a lightweight deep learning vision algorithm to support autonomous UAV landings and take-offs. It demonstrated near real-time performance on modern embedded GPU devices, and high safety and robustness in human presence detection and positioning error estimation. However, it’s yet to be tested in the field. 

Reliable Connectivity 

Autonomous UAVs with autonomous navigation systems will still require strong communication links with ground stations for localization and data exchanges (e.g., streaming video feeds, receiving updated flight plans, etc). Present-day communication ranges are limited to about 60 miles (35 km). Using lower frequencies allows higher ranges but limits the maximum data rate (i.e., results in higher latency). 

Software-defined networks (SDNs) for drones are emerging as a solution to this problem. With SDNs, the communication between the control layer and data link layer is commonly performed with OpenFlow protocol. OpenFlow protocol allows the network’s forwarding plane control over the network’s switch or router functionality. SDN design also improves the packet delivery ratio by tuning network speeds and UAV network security. 

The groundwork for enabling autonomous navigation systems for UAVs has already been substantial. In the coming years, we expect more solutions to move from the labs to the mainstream, ushering a new era of safe, autonomous aerial navigation. 

Machine learning and deep learning are staples in many industries. Ground transportation systems use predictive algorithms to optimize travel flows, predict congestion, and improve fuel usage. The sky’s the new limit. 

Impressive advances in AI navigation move from research labs to the mainstream, promising greater reliability, safety, and efficiency to aeronautics.

Inertial AI navigation for UAVs in GPS-compromised environments

GPS signal, necessary for most UAV operations, can be unavailable after a natural disaster or in military situations. Bavovna.ai, a graduate of the US Air Force Labs Mass Challenge acceleration program, is developing an AI-powered PNT navigation system for aerial, surface, and subsurface vehicles.

Designed for dual use, Bavovna’s inertial navigation system uses sensor fusion and pre-trained ML/DL algorithms to offer autonomous operations. Thanks to hardened core electronics, the system is resistant to common electromagnetic warfare threats. It’s also a low-SWaP solution with minimal power requirements and weight, suitable for UAV models and Class II aerial vehicles.

During trials, the Aurelia X6 Max multicopter could fly autonomously without remote control, GPS, or any other communication, collect location intelligence, and return to the home position. Bavovna AI navigation system aims for a 0.5% positioning error, even in complex journeys up to 30 miles (48 km). The team is working to extend the system use cases to cover Signals Intelligence, mine detection, automatic target engagement, and security surveillance.

AI copilots for commercial aviation flights

Modern aircraft have highly sophisticated autopilot systems to help pilots control altitude, course, thrust, and navigation. However, the onslaught of alerts and system interfaces requiring the pilot’s attention can be intimidating. NASA estimated that 34 different competing activities distract or preoccupy the pilots, ranging from communication to searching for VMC traffic. These can lead to human errors and dangerous consequences.

Air-Guardian, a new MIT CSAIL project, aims to improve the HMI of existing autopilot systems to ensure safer operations. The AI system uses eye-tracking to detect pilot distraction and “saliency maps” to understand aircraft behavior.

Based on a continuous-depth neural network model, the copiloting system can identify early signs of risks and take over the controls when needed. During trials, the Air Guardian system reduced flight risks and improved navigation success.

Advanced airflow traffic management to clear up congestion 

Unforeseen events like bad weather conditions adversely affect air traffic, causing congestion in a specific sector of the air navigation space. This, in turn, affects all network participants, leading to a cascade effect of delayed flights. 

Co-funded by the European Union and led by the Universita Ta Malta, the ASTRA project is short for AI-enabled tactical FMP hotspot prediction and resolution. The goal is to improve the prediction of air traffic congestion areas one hour in advance and suggest optimal resolution paths to traffic controllers.

The prediction algorithm will be trained on historical data (from 2019 to today) from the EUROCONTROL organization. The AI navigation system will provide FMPs with prescriptive scenarios to optimize flow management positions, ensuring safety, efficiency, lower fuel consumption, and environmental impacts.

AI has untapped potential in airborne navigation, and we expect more innovation in sensor fusion, AI-powered PNT, and aerial traffic management to happen in the next several years. 

GPS and INS are core components of global navigation systems, but both have inherent limitations. GPS is prone to signal blockage and jamming, as well as reflection on buildings and terrains, leading to signal distortion and poor position estimation. Signal accuracy also degrades at high speeds and rapid directional changes, making it less reliable for ultra-fast UAVs.

INS sensors are prone to drift due to poor initial calibration or sensor noise caused by mechanical vibration, temperature variations, cross-axis sensitivity, and wear. They also require periodic recalibration, especially for high-precision positioning.

Different machine learning and deep learning methods can improve GPS/INS signal reliability and quality—and, in some cases, even fully replace it.

The new era of Inertial navigation and AI GPS systems  

Artificial neural networks have already demonstrated strong results in real-time object recognition and signal processing in the automotive and manufacturing sectors. AI GPS in navigation systems are the next frontier.

Researchers tested different deep learning algorithms to optimize inertial sensing and sensor fusion tasks for land, aerial, and maritime vehicles. One group developed visual-inertial odometry (VIO), an integrated inertial and image sensor fusion technique based on deep learning to predict the UAS’s position with 0.167 Mean Square Error (RMSE).

Another group successfully used a convolutional neural network (CNN) to analyze noise information from the inertial measurement units using the Kalman filter, which improves the position and orientation estimation of the navigation system. Validation on real data from an autonomous underwater vehicle showed a 35.4% higher position accuracy than conventional methods.

AI GPS effectively compensates for individual sensor limitations by enabling rapid data processing from auxiliary sources. By leveraging an array of sensors and integrating AI algorithms, the vehicle can autonomously recalibrate, correct positional data, and operate smoothly in signal-compromised environments.

Advantages of AI-assisted navigation:

• Higher situational awareness and adaptability 
• Reduced dependence on satellite infrastructure 
• Lower vulnerability to signal jamming
• Better performance on complex terrain 
• Smarter collision detection and prevention 
• Improved coordination for drone swarms 

Real-world applications of AI sensor fusion for navigation 

At Bavovna, we’ve been developing a new AI-powered PNT navigation system to enhance existing inertial navigation systems of UAVs. By using AI for sensor fusion, Bavovna’s algorithms can safely pilot vehicles at high speeds and complex trajectories, even in GPS-compromised environments.

During a recent test, we successfully piloted a Radiobird Defender 001, a multi-weather VTOL, through its 30-km flying mission with an end-point positioning error of 0.42-0.63% on average.  

Sensor fusion from the onboard IMU array and airspeed provided navigation. We use sensors like a compass and barometer, with analysis performed on the edge using Bavovna AI Navigation Kit. We use a proprietary Kalman filtration algorithm and fine-tuned sensor fusion neural network trained on 100 hours of fight logs minimum.

SWaP efficient, the hardware kit is enclosed in a protective enclosure to minimize environmental and electromagnetic warfare impacts. The core electronics are also hardened with an EMI composite shield (-92 dBA/mm). 

AI GPS algorithms and sensor fusion provide a higher level of accuracy comparable to conventional GPS and GNSS, but with greater reliability and resilience. This results in more efficient drone operations due to enhanced autonomy, obstacle avoidance, and route optimization, even in GPS-denied environments.