Autonomous drones are increasingly used in cluttered, GPS-denied environments where safe and agile navigation depends on reliable visual obstacle avoidance. However, current approaches face three key challenges: the lack of a unified evaluation framework, the trade-off between safety and agility, and the gap between simulation-trained policies and real-world deployment.

This dissertation addresses these issues by developing learning-based methods and evaluation tools for onboard navigation. First, it introduces AvoidBench, a high-fidelity benchmarking suite with standardized environments and metrics to systematically evaluate obstacle avoidance performance.

Second, it presents MAVRL, a reinforcement learning algorithm that adapts flight speed to environmental complexity, achieving an improved balance between safety and agility. Third, it proposes Depth Transfer, a sim-to-real method that bridges differences in dynamics and perception, enabling robust deployment of trained policies on real drones.

Finally, a bio-inspired hierarchical architecture is introduced, separating high-level planning from low-level control to improve training efficiency and robustness.

Together, these contributions advance learning-based drone navigation by enabling reliable evaluation, adaptive behaviour, efficient training, and successful real-world deployment in complex environments. ... Read more

This work investigates adaptive switching mechanisms for reinforcement learning (RL) controllers in high-speed autonomous drone racing. While domain randomization (DR) improves generalization, single-policy controllers remain constrained by their training distributions and may fail under unmodeled conditions such as wind, hardware wear, or sensor degradation. To address this, we introduce two complementary online switching strategies: the look-back switch, which retrospectively identifies the dynamics model that best matches the current flight state, and the look-ahead switch, which prospectively selects the most suitable policy by simulating candidate controllers over short horizons. Both mechanisms leverage pre-trained policies without requiring retraining, enabling safer and faster flight. We evaluate the methods in simulation on two platforms (5-inch racing quadcopter and the high-fidelity A2RL competition drone) and across two tracks (Figure-8 and A2RL Grand Challenge). Results show that adaptive switching improves robustness and reduces lap times compared to fixed-policy baselines, demonstrating its potential for real-world deployment. ... Read more

We present the first demonstration of a fully spiking actor-critic neural network policy, trained via Proximal Policy Optimization (PPO), for continuous control of an agile high-speed quadcopter in a gate-based navigation task. The spiking neural network (SNN) controller employs Leaky Integrate-and-Fire neurons with surrogate gradient training and spike-rate decoding over multiple integration cycles, and it is benchmarked against a comparable artificial neural network (ANN) controller in both simulation and real-world flight tests. Results show that despite being trained to the same reward level, the SNN achieves superior performance in simulation, achieving higher episode rewards, greater robustness and reduced crash rate. Additionally, in 12-second real-world trials, the SNN outperforms the ANN, attaining a higher average reward (70.63 vs 59.77), greater mean velocity (7.94 vs 6.99 m/s), and more gates cleared (46.33 vs 40.67). An analysis of the spike integration cycle count reveals a clear trade-off: lower cycle counts (fewer integration steps per control update) reduce control output resolution and hinder learning, whereas higher cycle counts improve smoothness but increase inference latency. Moderate cycle counts (5 or 8) provide the best balance, yielding high rewards, smoother outputs, and low execution time overhead. These findings represent a key step forward for neuromorphic control in embedded autonomous systems, demonstrating that SNN-based policies can outperform conventional ANN controllers in high-speed, agile robotic tasks. ... Read more

In recent years, flapping wing micro aerial vehicles (FWMAVs) have garnered significant attention due to their agility in cluttered environments and the safety advantages offered by their soft wings during close-proximity operations. These vehicles are subject to numerous constraints: limited power supply, restricted payload capacity, high drag and vibration from the flapping mechanism, and susceptibility to environmental disturbances such as wind gusts—making indoor operation preferable. However, these vehicles have not achieved autonomous navigation yet. The platform used in this study, the Flapper Nimble+, is attitude-stable, meaning it lacks access to egomotion or positional information and is only capable of controlling its thrust, pitch, roll, and yaw. Given the strict SWaP (Size, Power and Weight) constraints, lightweight Time-of-Flight (ToF) sensors are among the most practical sensing solutions. This thesis investigates the question: How can obstacle avoidance be achieved in an attitude-stable flapping wing air vehicle equipped with Time-of-Flight sensors? Two control approaches were implemented and tested in the IsaacGym simulator: a simple PID controller with confidence-based yaw adjustment, and a reinforcement learning (RL) policy trained using Proximal Policy Optimization. Two sensor configurations were considered: a minimal two-sensor setup (front and downward) and a richer five-sensor array for broader perception. The PID controller successfully avoided collisions in all environments, demonstrating high reliability but limited exploration, averaging 35% coverage. In contrast, the RL policy with five sensors achieved greater spatial exploration—averaging 49.5% coverage—at the cost of an average of 3.5 collisions per episode. When reduced to two sensors, the RL agent’s performance declined significantly, with average coverage dropping to 29% and collision counts increasing to 6.6 (excluding failed runs). These results show that autonomous navigation is achievable for attitude-stable flapping wing air vehicles using only ToF sensors. While PID control offers reliable, conservative navigation under minimal sensing, RL enables more exploratory behaviors and adaptability in dynamic environments. ... Read more

The goal of this thesis was to push the boundaries of fast and agile flight for autonomous drone racing. A framework was developed for high-performance vision-based localization, sensor fusion, and control, with a focus on integration of neural networks. Additionally, a neural dynamics model was created which can can replace all accelerometer measurements by thrust&drag predictions, while running directly on the flight controller. ... Read more

The use of UAVs for aircraft maintenance inspections has emerged as a promising solution to reduce maintenance costs and increase inspection quality by providing efficient and accurate data collection. In this paper a novel path planning approach is proposed for autonomous aircraft maintenance inspections using UAV swarms, leveraging a frontier-following algorithm. The algorithm guides a drone's trajectory to cover an aircraft's surface while maintaining a constant distance tangent to the surface. Designed to operate with delayed depth sensing methods, where scenes are reconstructed using structure from motion principles, this approach can be integrated with online reconstruction techniques. The frontier-following methodology results in a characteristic spiraling pattern behavior and facilitates complex surface tracking. The primary environment sensing detector, an RGB camera, is angled towards a target on the frontier. This allows for preliminary mapping of a route prior to traversal, while a collision avoidance strategy ensures collision-free flight by tracking an obstacle-free tunnel around the drone. Furthermore, this method can be scaled up for application to swarming agents, accelerating surface coverage. Simulation results demonstrate the algorithm's efficiency in achieving comprehensive surface coverage. Frontier-following path planning outperforms the 2D coverage path planning algorithm Random Walk, while performing lower than Spiral Spanning Tree Coverage - a result consistent with the latter's optimality for 2D surfaces. The proposed method's ability to successfully track complex 3D surfaces, a capability lacking in traditional 2D coverage path planning algorithms, renders it a promising solution for autonomous inspection systems. ... Read more

There exists a wide array of possible applications for small, safe, cost-effective, and energy-efficient drones.
However, their development is hampered by limited payload capacity, which restricts both computational power and flight time.
Traditional control systems and sensor processing algorithms are ill-suited for these resource-constrained platforms since they typically rely on power-hungry processors and complex numerical methods.

This thesis investigates neuromorphic approaches to both state estimation and control for small drones.
Inspired by the energy-efficient and highly parallel processing of biological neural systems, neuromorphic computing leverages spiking neural networks (SNNs) that operate via discrete spikes, offering real-time, low-power processing capabilities for micro aerial vehicles (MAVs).
While previous work has applied neuromorphic methods to high-level perception tasks, their application to fundamental flight control -- such as precise attitude estimation and low-level control -- remains largely unexplored.

Following a review of the current state of neuromorphic computing, the research first explores its application to state estimation.
A recurrent SNN is designed to estimate the drone’s attitude from inertial measurement unit (IMU) data, achieving performance comparable to conventional methods like the complementary filter, despite employing a minimal network architecture.
The study then investigates event-based vision sensors by processing data from a downward-facing event camera to estimate the attitude and angular rates, enabling a quadrotor to achieve flight without inertial sensing -- a pioneering demonstration in the field.

Transitioning from estimation to control, the thesis uses neuromorphic algorithms to perform low-level control tasks.
A spiking PID controller is developed using a fixed network architecture, demonstrating altitude control using Intel's Loihi neuromorphic processor.
To address the challenge of precise integration inherent in spiking systems, the Input-Weighted Threshold Adaptation (IWTA) mechanism is introduced.
This innovative approach allows for precise integration of incoming signals and was used as the integral component of a neuromorphic PID controller, mitigating steady-state errors and compensating for sensor biases.

Ultimately, the work unifies estimation and control into a single end-to-end neuromorphic system deployed on a tiny 27g Crazyflie quadrotor. Trained via imitation learning on real flight data, the integrated network maps raw inertial sensor inputs directly to motor commands at a control frequency of 500Hz, achieving attitude tracking performance comparable to traditional controllers.

Overall, this thesis demonstrates that neuromorphic computing is a promising approach for low-level state estimation and control in flying drones, while also addressing the challenges of implementing such systems in real-world environments with sensor biases and persistent disturbances. ... Read more

UAVs have become increasingly popular, finding applications in diverse areas such as military operations, search and rescue, delivery services, wireless communication, and aerial surveillance. A critical aspect of UAV operations is autonomous landing, especially on moving targets like ships, which presents significant challenges due to the dynamic and unpredictable maritime environment. This research explores the potential of reinforcement learning as a strategy for achieving optimal guidance during the autonomous landing process of the VSQP. The study illustrates that the reinforcement learning learning framework can effectively steer the VSQP, ensuring safe and accurate landings on a moving ship, and outperforming the benchmark controller in a more intelligent manner. The proposed approach not only improves landing performance but also extends the operational capability of the VSQP. ... Read more

Quadrotors have continuously leveraged the use of artificial intelligence for navigation and decision-making. Moreover, neuromorphic computing, specifically Spiking Neural Networks (SNNs), is considered as an energy-efficient solution during inference. The current study will analyse the effects of implementing SNNs for mimicking energy optimal guidance and control. To achieve this, population encoding is used and an equivalent of 7-8 spiking neurons per conventional neuron is found to preserve most of the information. The equivalent controller prefers fast adaptation which requires small spiking threshold values and minimal reliance on past information. To improve the controller performance, dataset selection is of utmost importance with a careful trade-off between excessive race track customisation and generalisability being required. The results show that learning is feasible and SNN performance approaches conventional state-of-the-art models trained with multi-layer perceptrons. The current analysis represent an important step towards the rapid guidance and control of ultra-small energy efficient quadrotors. ... Read more

Autonomous drone racing has gained attention for its potential to push the boundaries of drone navigation technologies. While much of the existing research focuses on racing in obstacle-free environments, few studies have addressed the complexities of obstacle-aware racing, and approaches presented in these studies often suffer from overfitting, with learned policies generalizing poorly to new environments. This work addresses the challenge of developing a generalizable obstacle-aware drone racing policy using deep reinforcement learning. We propose applying domain randomization on racing tracks and obstacle configurations before every rollout, combined with parallel experience collection in randomized environments to achieve the goal. The proposed randomization strategy is shown to be effective through simulated experiments where drones reach speeds of up to 70 km/h, racing in unseen cluttered environments. This study serves as a stepping stone toward learning robust policies for obstacle-aware drone racing and general-purpose drone navigation in cluttered environments. Code is available at https://github.com/ErcBunny/IsaacGymEnvs. ... Read more

In recent years, the use of drones in practical applications has seen a rapid increase, for instance in inspection, agriculture or environmental research. Most of these drones have a span in the order of tens of centimeters and a weight of half a kilogram or more. Smaller drones offer advantages in terms of safety and cost. However, their reduced payload capacity makes it difficult to carry the sensors and computers required for autonomous operation.

One of the most essential tasks an autonomous drone needs to perform is navigation. Here, navigation is defined as the ability to move towards a specified location while avoiding obstacles along the way. Ideally, the drone should also remember traveled routes, to make the return journey more efficient. However, on tiny drones (palm-size or smaller) the on-board processing power is often limited to a single microcontroller and the choice of sensors is limited. Cameras are popular sensors for tiny drones, because they're small, lightweight and passive, although they do require some processing power to produce useful results. The goal of this dissertation is to find a new, visual navigation strategy that fits within the constraints of these tiny drones.

First, existing work in terms of visual perception and avoidance is reviewed. Multiple options exist for visual perception: stereo vision, optical flow and monocular vision. All of these options are discussed and compared, leading to the conclusion that stereo vision performs best at shorter distances albeit at the cost of an additional camera, while monocular vision performs better at longer distances. Optical flow is ruled out for avoidance, as it has excessively large errors precisely in the direction of movement.
For avoidance, the options in terms of motion planning, map types and odometry are discussed. Perhaps unsurprisingly, the optimal choice is found to be dependent on the application. For computational efficiency on tiny drones, the most important choice is whether multiple measurements should be fused into a single map, or if individual percepts are good enough for avoidance. The latter is significantly less computationally demanding. For visual odometry, the depth information should be used if available, and the IMU can provide efficiency benefits in feature tracking. At Preliminary results are shown for monocular vision, visual odometry and obstacle avoidance.

Secondly, the dissertation takes a deeper dive into monocular depth estimation. Monocular depth estimation has the advantage that it only needs a single camera -- which saves valuable weight on tiny drones -- but its processing is more complex. The goal of this chapter is to analyze the learned behavior of neural networks for monocular depth perception, to see if this can be distilled into simple, lightweight algorithms. Using experiments based on data augmentation, it is shown that all four of the analyzed networks rely on the vertical position of objects in the image to estimate their depth. While this cue would be simple to replicate, it does depend on a known pose of the camera. Further investigation shows that the networks have a strong prior `assumption' about this pose, which may make transfer to drones more difficult. Finally, the networks need to have some sense of an `object'. In this case, it is shown that various shapes are recognized as an object provided that they have contrasting outlines and a dark shadow at the bottom. While this last feature is clearly present in the car-based KITTI dataset, it may not transfer directly to other environments. However, the vertical position cue can likely be used to provide monocular depth estimates to resource-limited systems such as tiny drones.

Thirdly, the remembering of traveled routes is investigated. Traditional mapping strategies from robotics would quickly run out of memory on microcontrollers, especially over longer trajectories. Instead, inspiration for a memory-efficient route-following strategy is found in nature. Here, insects are able to remember and follow remarkably long routes despite their tiny brains. Their strategy is often broken up into a few components, most notably path integration (odometry in robotics) and visual homing. We implement a novel strategy based on these components on a 56-gram drone. Here, the focus lies on traveling long distances using odometry, while periodically using visual homing to return to known locations to counteract odometric drift. The proposed strategy is demonstrated over multiple experiments, where the most efficient run required only 0.65 kilobytes to remember a route of 56 meters. This shows that tiny drones can retrace known paths by combining odometry with periodic homing maneuvers to counteract drift.

Finally, the avoidance of obstacles is discussed in the conclusion of this dissertation. This research has been performed by MSc students under my supervision, who have found and demonstrated that bug algorithms are an effective navigation strategy in three-dimensional, limited-field-of-view applications and provide a lightweight goal-oriented avoidance strategy that is suitable for tiny drones.

By combining all of the above results, a full navigation strategy for tiny drones can be proposed: tiny drones can visually navigate by using lightweight monocular vision algorithms to perceive obstacles, three-dimensional bug algorithms to avoid them while moving to new locations, and odometry and visual homing to retrace known paths. ... Read more

With their ability to access hard-to-reach spaces and to provide a birds-eye view over large areas,Micro Air Vehicles (MAV) are already taking over awide variety of monitoring and inspection tasks. As the continuing miniaturization of electronic components further drives down cost, recent advancements in the design of control algorithms promise a future, where these small drones will be able to operate fully autonomously and with minimal human input. Spatial awareness is a key aspect in developing fully autonomous MAVs and in this dissertation,we develop new algorithms for on-board localization usingUltra-Wideband (UWB) ranging. UWB is a low-power technology that enables data communication and time-of-flight range measurements. The large bandwidth of UWB results in high timing accuracy as well as an improved resistance to shadowing and multipathing, which makes UWB a very useful technology for ranging in indoor environments... ... Read more

Executing quadrotor trajectories accurately and therefore safely is a challenging task. State-of-the-art adaptive controllers achieve impressive trajectory tracking results with slight performance degradation in varying winds or payloads, but at the cost of computational complexity. Requiring additional embedded computers onboard, adding weight and requiring power. Given the limited computational resources onboard, a trade-off between accuracy and complexity must be considered. To this end, we implement "Neural-Fly" a lightweight adaptive neural controller to adapt to propeller damage, a common occurrence in real-world flight. The adaptive neural architecture consists of two components: (I) offline learning of a condition invariant representation of the aerodynamic forces through Deep Neural Networks (II) fast online adaptation to the current propeller condition using a composite adaptation law. We deploy this flight controller fully onboard the flight controller of the Parrot Bebop 1,showcasing its computational efficiency. The adaptive neural controller improves tracking performance by ≈60% over the nonlinear baseline, with minimal performance degradation of just ≈20% with increasing propeller damage. ... Read more

Reaching fast and autonomous flight requires computationally efficient and robust algorithms. To this end, we train Guidance & Control Networks to approximate optimal control policies ranging from energy-optimal to time-optimal flight. We show that the policies become more difficult to learn the closer we get to the time-optimal ’bang-bang’ control profile. We also assess the importance of knowing the maximum angular rotor velocity of the quadcopter and show that over- or underestimating this limit leads to less robust flight. We propose an algorithm to identify the current maximum angular rotor velocity onboard and a network that adapts its policy based on the identified limit. Finally, we extend previous work on Guidance & Control Networks by learning to take consecutive waypoints into account. We fly a 4×3m track in similar lap times as the differential-flatness-based minimum snap benchmark controller while benefiting from the flexibility that Guidance & Control Networks offer. ... Read more

Aerial platforms designed for water jet placement are gaining interest in the areas of fire-fighting, washing, and irrigation. A novel, lightweight, and simplistic design is proposed that reduces the number of actuators and limits ineffective water discharge. External camera feedback was used for position control as a first step towards autonomous flight. An initial prototype of an unmanned hydro-propelled aerial vehicle (UHAV) connected to a water hose was designed and fabricated. Flight tests were conducted to show that attitude control with uniaxial thrust-vectoring of two nozzles was impossible due to undamped vibrations and coupling effects. By redesigning the PID controller, pitch rate damping was accomplished. Furthermore, a design trade-off led to the introduction of a canting keel to reduce bank-yaw coupling effects due to asymmetric nozzle deflections. Flight tests proved that the iterated design with a hose length of 3m was capable of disturbance rejection and setpoint tracking. An external camera was used to show that the Lucas-Kanade optical flow algorithm and the implementation of the YOLOv5 segmentation model can be used for positional water jet placement. By increasing the pitch rate damping, improving the water jet detection algorithm and implementing a cost function for water discharge at the area of interest, autonomous missions can be flown in the future. ... Read more

In this study, we present a first step towards a cutting-edge software framework that will enable autonomous racing capabilities for nano drones. Through the integration of neural networks tailored for real-time operation on resource-constrained devices. A lightweight Convolutional Neural Network, with the Gatenet architecture, is adjusted for reduced computational demand and is successfully deployed on a GAP8 processor at a rate of 16$Hz$. This network provides gates' size and location data for the subsequent positioning algorithm. A second neural network, trained through reinforcement learning, governs the drone's guidance and control systems, demonstrating a remarkable rate of 167$Hz$ on an STM32F405 processor. The attitude rates and thrust outputted by this network are then fed to an attitude rate PID controller.

The research shows that state-of-the-art neural networks for drone racing can be deployed on nano drones, despite their limited processing power. Nonetheless, the study demonstrated specific limitations, such as the perception network's sensitivity to white pixels in the image reducing its effectiveness when light sources are present in the scene. These findings underscore the importance of dataset composition and the need for diverse training scenarios to enhance the neural network's generalizability and performance in real-world applications. ... Read more

In the ever-evolving landscape of robotics, the quest for advanced synthetic machines that seamlessly integrate with human lives and society becomes increasingly paramount. At the heart of this pursuit lies the intrinsic need for these machines to perceive, understand, and navigate their surroundings autonomously. Among the senses, vision emerges as a cornerstone of human perception, providing a wealth of information about the world we inhabit. Thus, it comes as no surprise that equipping robots with vision-based perception capabilities, or computer vision, has captivated researchers for decades. Recent breakthroughs, fueled by the advent of deep learning, have propelled computer vision to new heights. However, challenges persist in leveraging the power of deep learning, as its hunger for computational resources poses hurdles in the realm of robotics, particularly for small flying robots with their inherent limitations of payload and power consumption.

This dissertation embarks on a journey that begins at the intersection of two groundbreaking technologies with the potential to revolutionize computer vision and enhance its accessibility to small robots: event-based cameras and neuromorphic processors. These two technologies draw inspiration from the information processing mechanisms employed by biological brains. Event-based cameras output sparse events encoding pixel-level brightness changes at microsecond resolution, while neuromorphic processors leverage the power of spiking neural networks to realize a sparse and asynchronous processing pipeline.

Throughout this dissertation, comprehensive investigations have been conducted, presenting innovative solutions and advancements in the fields of computer vision and robotics. The thesis begins by presenting the winning solution of the 2019 AIRR autonomous drone racing competition, which showcases a monocular vision-based navigation approach specifically designed to address the limitations of conventional sensing and processing methods. Moreover, it explores the bridging of the gap between event-based and framebased domains, enabling the application of conventional computer vision algorithms on event-camera data. Building upon these achievements, the thesis introduces a pioneering spiking architecture that enables the estimation of event-based optical flow, with emergent selectivity to local and global motion through unsupervised learning. Additionally, the thesis presents a framework that addresses the practicality and deployability of the models by training spiking neural networks to estimate low-latency, event-based optical flow with self-supervised learning. Finally, this dissertation culminates with a demonstration of the integration of neuromorphic computing in autonomous flight. By utilizing an eventbased camera and neuromorphic processor in the control loop of a small flying robot for optical-flow-based navigation, this research showcases the practical implementation of neuromorphic systems in real-world scenarios. Overall, our studies demonstrate the benefits of incorporating neuromorphic technology into the vision-based state estimation pipeline of autonomous flying robots, paving the way for the development of more power-efficient and faster neuromorphic vision systems. ... Read more

Developing optimal controllers for aggressive high speed quadcopter flight remains a major challenge in the field of robotics. Recent work has shown that neural networks trained with supervised learning are a good candidate for real-time optimal quadcopter control. In these methods, the networks (termed G\&CNets) are trained using optimal trajectories obtained from a dynamical model of the quadcopter by means of a direct transcription method. A major problem with these methods is the effects of unmodeled dynamics. In this work we identify these effects for G\&CNets trained for power optimal full state-to-rpm feedback. We propose an adaptive control strategy to mitigate the effects of unmodeled roll, pitch and yaw moments. Our method works by generating optimal trajectories with constant external moments added to the model and training a network to learn the policy that maps state and external moments to the corresponding optimal rpm command. We demonstrate the effectiveness of our method by performing power-optimal hover-to-hover flights with and without moment feedback. The flight tests show that the inclusion of this moment feedback significantly improves the controller's performance. Additionally we compare the adaptive controller's performance to a time optimal Bang-Bang controller for consecutive waypoint flight and show significantly faster lap times on a 3x4m track. ... Read more

Flying-wings show great potential for a vast number of applications, in both commercial and military sectors, thanks to their long range and fast forward flight, but suffer due to their lack of vertical take-off and landing capabilities. This paper presents a proof of concept for a novel landing method for a conventional flying wing that does not introduce additional weight dedicated only to the landing phase, with the aim of controlling a deep-stalled flying-wing in a powered flat spin. Through cyclic actuation of the servo motors and elevons, lateral forces as well as moments can be generated to control the position and attitude of the rotation plane. A successful indoor experiment was performed with a modified Parrot Disco in a controlled environment. Outdoor tests, however, failed to replicate the indoor results due to additional challenges present in the real flight conditions. A number of key challenges were identified, and the insights gained in this research lay an initial foundation for future work on this topic. ... Read more