cs.ROOct 27, 2026

Autonomous fixed-wing interception using reinforcement learning

Authors: Daan van Hamersveld, Ewoud Smeur, Iurii Kriachko

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Medium altitude, long endurance unmanned aerial systems such as Shahed-type strike drones are an unanswered threat in modern conflicts. Conventional air defense systems are too expensive to provide a feasible large-scale solution. Quadcopter interceptors are a cheap, battle-proven alternative, but their extremely limited flight time restricts them to low altitude targets at short range, whereas fixed-wing platforms have the endurance to reach higher altitude targets from longer distances. However, such interceptors are still operated manually. Recent developments in learning-based controllers show potential for automation. In this work, a fixed-wing interceptor is presented that is controlled by a neural network that maps filtered monocular vision observations directly to actuator commands. The policy was trained in a custom Python environment using reinforcement learning and proximal policy optimization. The policy was then deployed on a fixed-wing platform and tested in real interception scenarios. As a baseline, two classical guidance controllers were also developed and tested in simulation. The policy is able to intercept targets both in simulation and in real flight. In simulation, the policy outperforms the classical controllers under most conditions. The interceptor was ultimately successfully validated in real flight by intercepting maneuvering fixed-wing targets. This novel system shows the potential of artificial intelligence for autonomous control of fixed-wing interceptor drones.

Figures & tables

Fig. 1: The interceptor during a high bank angle.
Fig. 2: A block diagram showing the constructed feedback loops that power the baseline classical autopilot. In simulation, sideslip is computed as β=arcsin⁡ ⁣(v/∥vb∥)\beta=\arcsin\!\left(v/{\lVert\mathbf{v}_{b}\rVert}\right) , with vv the lateral velocity in the body frame. In real flight, Ardupilot’s EKF estimate for sideslip can be used. Note that unlike the baseline controllers, the RL policy does not observe β\beta .
CategoryCoefficientFlight 1Flight 2Flight 3
Pitching momentR2\mathbf{R}^{2}0.47940.68090.6889
Cm0C_{m0}-0.0035-0.0113-0.0127
CmαC_{m_{\alpha}}-0.0508-0.1885-0.1724
CmqC_{m_{q}}-1.9179-3.0066-2.4160
CmδeC_{m_{\delta_{e}}}0.15360.36470.2907
LiftR2\mathbf{R}^{2}0.59430.78530.7967
TABLE I: Estimated Aerodynamic Stability and Control Derivatives
Fig. 3: A flight by flight comparison of the estimated aerodynamic coefficients. The dashed line indicates identical estimates between two flights.
Fig. 4: A ten second segment from flight 2 showcasing how the model identified using flight 3 (orange line) compares to the measured aerodynamic coefficients from flight 2 (blue line) over time.
Fig. 5: The autocorrelation function of the residuals for each aerodynamic coefficient.
Fig. 6: The performance of the cascaded autopilot controller in the full non-linear model. (a) The inner roll-rate loop tracking response to a 0.5, 1, 2 and 4 rad/s roll rate demand. (b) The inner pitch-rate loop tracking response to a 0.1, 0.2, 0.4 and 0.8 rad/s pitch rate demand. (c) The inner sideslip angle loop nullifying response to a 5, 10, 15 and 20 degree initial sideslip. (d) The outer roll angle tracking loop response to a 15, 30, 45 and 60 degree roll demand. (e) The outer normal acceleration tracking loop response to a 3, 6, 10 and 15 m/s 2 normal acceleration demand.
Fig. 7: The top and bottom figures showcase the progression of the average interception rate and reward throughout the curriculum respectively. Data is plotted in increments of a thousand episodes over the cumulative amount of training steps. Each colour represents a separate stage.
Fig. 8: The trajectories of the interceptor (blue) and target (red) throughout the course of a single simulated engagement. The circle and cross indicate the start and finish respectively. The dashed orange line indicates the KF estimate of the target’s position.
Fig. 9: The ground truth (GT) range (blue) versus the Kalman Filter (KF) range estimate (dashed orange). The red areas indicate periods where the target is not being detected. The dashed green line indicates the range threshold for interception.
Fig. 10: The angle of attack throughout the simulated engagement. The dashed lines indicate the angle thresholds beyond which high absolute angle of attack is punished. The red dotted line indicates the cutoff threshold to end the episode early due to stall or airframe collapse.
Fig. 11: The state of the actuators throughout the simulated engagement. Aileron, rudder and elevator deflection are represented by the red, green and blue lines respectively. Throttle percentage [0,1] is shown by the brown line.
Fig. 12: The interceptor’s attitude in Euler angles throughout the simulated engagement. Roll, pitch and yaw are indicated by red, green and blue lines respectively.
Fig. 13: The trajectories of the RL interceptor (blue), PN interceptor (orange), PP interceptor (green) and target (red) throughout the course of a single engagement. The circle and cross indicate the start and finish respectively. Noise and other randomness was seeded identically for each interceptor.
Fig. 14: The Euler angles of the classically guided interceptors during the same simulated engagement. Roll, pitch and yaw are represented in red, green and blue respectively. The top and bottom subfigures shows the PN and PP interceptors respectively.
Fig. 15: The control surfaces of the classically guided interceptors during the same simulated engagement. Elevator, aileron and rudder are represented in blue, red and green respectively. The top and bottom subfigures shows the PN and PP interceptors respectively. Throttle is set to maximum at the start of the engagement for both controllers.
Fig. 16: The performance of the RL (blue), PN (orange) and PP (green) controllers throughout the nine stages. The solid lines and circles represent performance with a 60 ∘ FOV, noisy camera undergoing increasing disturbances as specified in appendix IV . The dashed lines and triangles represent performance with an ideal camera with a 180 ∘ FOV, with perfect detections at all stages. Each point represents 896 simulated episodes (64 parallel environments running 14 episodes each). The interception rate plotted on the y-axis represents the fraction of those episodes ending with a successful interception. The shaded area represents the 95 percent confidence interval.
Fig. 17: Interception rates for increasingly difficult conditions. The error bars indicate the Wilson 95 percent intervals. The nominal conditions of stage 5 hold except for one variable. In (a), the interception rate is shown for increasingly faster, non-accelerating targets. In (b), the target is increasingly accelerating with a mean initial speed of 7.3 m/s. In (c), the max wind speed increases up to 9 m/s. In (d), the mean detection dropout length increases up to 2.8 s. In (e), the range is increased. In (f), the approach geometry is varied from a tail-chase to a head-on engagement.
Fig. 18: A breakdown of the reasons episodes were terminated, categorized by condition. Green, purple, orange, red and grey represent the fraction of episodes terminated due to successful interception, target loss, near misses, loss of control and timeout.
Fig. 19: 3D trajectories of the interceptor during 6 real flight engagements of a stationary virtual target. The green dot represents the start of the engagement, that is the interceptor is flying in MANUAL mode and the RL controller is overriding RC. The red star indicates the relative position and altitude of the virtual target. The black cross indicates the closest point of approach (CPA) in meters. The viridian color scheme shows how the trajectory progresses through time. The aircraft in white with grey outline is a schematic representation of the interceptor’s attitude at discrete time steps. The aircraft model is not to scale.
Fig. 20: The commanded flight surface deflections and throttle during each real flight engagement of a virtual stationary target. In blue, orange and green, the normalized surface deflection of the ailerons, elevator and rudder respectively. The dashed grey line represents the commanded throttle.
Fig. 21: 3D trajectories of the interceptor during 8 real flight engagements of a moving virtual target. The green and red dot represent the interceptor’s and target’s position respectively at the start of the engagement. The red and black cross indicate the closest point of approach (CPA) of the target and interceptor respectively. The grey dashed line shows connects the target and interceptor at the CPA. The red line shows the target’s trajectory. The viridian color scheme shows how the interceptor’s trajectory progresses through time. The aircraft in white with grey outline is a schematic representation of the interceptor’s attitude at discrete time steps. The aircraft model is not to scale.
Fig. 22: The distance between the interceptor and the target throughout the course of the eight engagements. Each colour indicates a separate engagement, with the cross in the corresponding colour representing the shortest distance between the interceptor and the target. CPA: closest point of approach.
Fig. 23: 3D trajectories of the interceptor during 6 real flight engagements of a fixed-wing target. The green and red dot represent the interceptor’s and target’s position respectively at the start of the engagement. The red and black cross indicate the closest point of approach (CPA) of the target and interceptor respectively. The grey dashed line shows connects the target and interceptor at the CPA. The red line shows the target’s mission flight path. The CPA here is computed based on the proximity to the mission flight path, not the actual target trajectory. The viridian color scheme shows how the interceptor’s trajectory progresses through time. The aircraft in white with grey outline is a schematic representation of the interceptor’s attitude at discrete time steps. The aircraft model is not to scale. The target’s trajectory shown here is an estimated trajectory, assuming a constant 15 m/s speed and no maneuvering or deviation from the predetermined flight path. Actual CPA may have been smaller because the target was also maneuvering which is not represented in this figure. The target trajectory is computed based on the assumption that the target had a constant speed and the red line is simply an extrapolation of where the target would have been if the theoretical CPA was indeed the closest point of approach.
Fig. 24: Engagement 3. 3.3 meter range estimate. The date and time stamp are incorrect and can be neglected; they are only included to show the full frame.
Fig. 25: Engagement 5. 6.0 meter range considering approximately 45 degrees of yaw with respect to the optical axis of the interceptor. The date and time stamp are incorrect and can be neglected; they are only included to show the full frame.
Fig. 26: Engagement 6. 8.1 meter range considering approximately 20 degrees of yaw with respect to the optical axis of the interceptor. The date and time stamp are incorrect and can be neglected; they are only included to show the full frame.
Fig. 27: 1 Hz KF logs of the estimated range (a), position estimate variance (b), estimated target velocity (c), velocity estimate variance (d), time elapsed since the last detection (e). Each colour represents a single engagement plotted on a common time axis.
Fig. 28: The root mean square (rms) in degrees of rudder deflection versus the range between the target and the interceptor. The blue line shows rudder deflection amplitude against real range as seen when intercepting stationary targets. The purple and yellow dashed lines show the rudder deflection amplitude at various ranges based on presented artificial range estimates.
Appendix figures & tables4 assets

Supplementary material from the paper’s appendix.

Appendix

SymbolValueUnitDescription
Roll-rate loop (inner, aileron)
Kff,pK_{ff,p}0.134s rad -1rate feedforward, 1/pss1/p_{ss}
KP,pK_{P,p}0.25s rad -1proportional gain
KI,pK_{I,p}0.15rad -1integral gain
Pitch-rate loop (inner, elevator)
Kff,qK_{ff,q}0.300s rad -1rate feedforward ( 1/qss=1/q_{ss}= 0.405)
Appendix
TABLE II: Gains and constants of the cascaded acceleration autopilot of figure 2 . Surface commands are normalised to [−1,1][-1,1] , so gains acting on a surface are dimensionless.
SymbolDescriptionValue
Term weights
cdc_{d}Distance-progress reward1.01.0
cintc_{\text{int}}Interception bonus0.50.5
ccc_{c}Camera-centering reward0.40.4
cΦc_{\Phi}Field-of-view penalty1.51.5
cΩc_{\Omega}Body-rate penalty0.70.7
Appendix
TABLE III: Coefficients and parameters of the reward function.
CategoryCoefficientValue
LiftCL0C_{L0}0.31
CLαC_{L_{\alpha}}2.90
CLδeC_{L_{\delta_{e}}}0.50
DragCD0C_{D0}0.12
kinducedk_{\mathrm{induced}}0.18
Side ForceCYβC_{Y_{\beta}}-0.096244
Appendix
TABLE V: Aerodynamic Stability and Control Derivatives
Fig. 29: A breakdown of the reasons episodes were terminated, categorized by condition. Green, purple, orange, red and grey represent the fraction of episodes terminated due to successful interception, target loss, near misses, loss of control and timeout. All conditions except for nominal are in some way beyond the training envelope, giving insight into the robustness of the policy. These variations were measured against stage 8 nominal conditions instead of stage 5 nominal conditions as in in figure 18 .
More in this topic

Aerial Robotics

Aerial robotics studies the design, perception, control, and coordination of flying robots for navigation, inspection, transport, and physical interaction.

See all 340 papers

Explore similar work

CardsList
Mar 17, 2026

Agile Interception of a Flying Target using Competitive Reinforcement Learning

This article presents a solution to intercept an agile drone by another agile drone carrying a catching net. We formulate the interception as a Competitive Reinforcement Learning problem, where the interceptor and the target drone are controlled by separate policies trained with Proximal Policy Optimization (PPO). We introduce a high-fidelity simulation environment that integrates a realistic quadrotor dynamics model and a low-level control architecture implemented in JAX, which allows for fast parallelized execution on GPUs. We train the agents using low-level control, collective thrust and body rates, to achieve agile flights both for the interceptor and the target. We compare the performance of the trained policies in terms of catch rate, time to catch, and crash rate, against common heuristic baselines and show that our solution outperforms these baselines for interception of agile targets. Finally, we demonstrate the performance of the trained policies in a scaled real-world scenario using agile drones inside an indoor flight arena.

Timothée Gavin, Simon Lacroix, Murat Bronz

ENAC-LAB, LAAS-RIS · IAS, Thales LAS, Rungis, France · LAAS-RIS +3

Robotic RL
AskPDF
Jul 2, 2026

Learning Agile Intruder Interception using Differentiable Quadrotor Dynamics

This paper presents a methodology for learning a control policy to intercept an intruder using the 3D direction unit vector to the intruder and the interceptor state. Prior deep reinforcement learning approaches assume either relative position or distance to the intruder is available, but this information is not readily accessible in real-world applications that employ passive, monocular camera sensors. Instead, we propose a solution that leverages an analytical policy gradient method using differentiable quadrotor dynamics to learn agile interception at speeds up to 10 m/s. The proposed approach outperforms baseline methods that utilize simplified point mass dynamics by an average of 30%.

Michael Anoruo, Xiaoyu Tian, Abhishek Rathod +5

Carnegie Mellon University · Artificial Intelligence Integration Center

RL ControlRobot Policy LearningQuadrotor ControlPolicy Learning
AskPDF
Oct 26, 2026

ISB-NMPC: Interception-Oriented Spherical-Bearing Nonlinear Model Predictive Control for Quadrotors

Quadrotor interceptors equipped with strapdown cameras offer a low-cost solution for physically countering unauthorized UAVs. However, this configuration inherently couples vehicle maneuvering and camera pointing, creating a conflict between rapid closing and maintaining target visibility. We present interception-oriented spherical-bearing nonlinear model predictive control (ISB-NMPC), which couples spherical-bearing prediction with nonlinear quadrotor dynamics to jointly optimize target-directed closing motion and visibility. An attitude-compensated extended Kalman filter on S^2 estimates target bearing motion, compensates for perception delay, and predicts target bearings to guide scale-adaptive detection. Across 30 paired simulation trials, ISB-NMPC achieves an interception success rate of 86.7%, the highest among the evaluated controllers, with a mean per-trial target-visible time of 99.11%. Outdoor flights demonstrate acquisition of a 30 cm target at approximately 150 m, with interceptor speed reaching 17 m/s in the long-range flight and approximately 14 m/s while pursuing a quadrotor-towed maneuvering target. The software and hardware are open source at https://github.com/HOLYNOTBAD/ISB-NMPC and https://github.com/HOLYNOTBAD/holydrone, respectively.

Li Hong, Linkai Liu, Xiangyu Ren +6

Li Hong, Chen Min, and Yujun Zhang are with Shen Yuan Honors College, Beihang University, Beijing, China · Linkai Liu, Xiangyu Ren, Shuli Lv, Dingchen Zeng, and Quan Quan are with the School of Automation Science and Electrical Engineering, Beihang University, Beijing, China. · Chunjiang Wu is with the Hangzhou International Innovation Institute, Beihang University, Hangzhou, China.

Quadrotor ControlNonlinear MPC
AskPDF