MSc Advanced Aerospace Engineering

Control of a Fixed-Wing
Multibody UAV

An interactive companion to the dissertation — the hinged multibody aircraft, its instability, and the LQR / LQG control that tames it. Every rig below runs live control math in your browser.
AuthorSiddarth Jawahar
SupervisorDr Yinan Wang
InstitutionUniversity of Liverpool
ProjectNo. 213 · Sept 2023
0
Live control rigs
0
Control laws · LQR / LQG
0
Hinge sweep
0
Math solved in-browser
Scroll to explore
01
01 The Concept

One wing, many bodies

A multibody UAV joins several small aircraft wingtip-to-wingtip through hinges, forming a single high-aspect-ratio wing. It keeps the efficiency of a long, slender span while retaining the strength and redundancy of independent aircraft — if one body fails, the others hold the formation. The catch: connect them and the whole assembly becomes aerodynamically unstable, and it cannot fly without a deliberately designed control system.

High aspect ratio

A long, jointed span gives a high lift-to-drag ratio and long endurance — inspired by the X-HALE flexible test aircraft.

Hinged coupling

Wingtip hinges pass roll between bodies but block bending moments — cutting structural weight while adding coupled dynamics.

Redundant & agile

Each body can hold its own controller, take off and land independently, and reconfigure the wing shape in flight.

The engineering question this work answers: is it possible to keep such an aircraft stable when every added body makes the lateral modes worse — using nothing but classical control fundamentals applied a smarter way? The journey below builds up from the simplest unstable system to the full variable-geometry twin.

02
02 Control Theory

The toolkit: state space, LQR & LQG

Every model here is written in state-space form and stabilised with an optimal regulator. These are the exact tools used throughout the dissertation — and the exact math running in the rigs below.

State space
ẋ = Ax + Bu + Ew
y = Cx + Du + Fw
x — state vector (roll, rates, …) · u — control (aileron, elevator, rudder, thrust) · w — gust disturbance · A,B,C,D — system matrices
Linear Quadratic Regulator
J = ∫0 (xTQx + uTRu) dt
u = −Kx,  K = R−1BTP
P solves the algebraic Riccati equation: ATP + PA − PBR−1BTP + Q = 0. Q rewards holding the state; R penalises control effort.
Linear Quadratic Gaussian
x̂ = A + Bu + L(y − C)
LQR assumes you can measure every state. LQG pairs the regulator with a Kalman filter that estimates the state from noisy sensors, then feeds the estimate to K. Optimal control meets optimal estimation.
Stability
Re( λ(A) ) < 0  ⇒  stable
The eigenvalues λ of A are the flight modes (phugoid, Dutch roll, spiral, roll…). A single eigenvalue in the right half-plane means the aircraft diverges. Control moves the poles left.

Fidelity note — the inverted-pendulum rig uses the exact A, B matrices from the dissertation appendix and solves the Riccati equation live. The aircraft rigs use representative lateral-dynamics models tuned to reproduce the report's roll-response and eigenvalue behaviour, since the original state-space .mat matrices are not embedded here.

03
03 Rig I · The Primer

Inverted pendulum — why LQR

The classic unstable system: a pole balanced above a cart. Like the aircraft, it senses its own tilt and moves to stay upright. Drag the cart or nudge the pole — the live LQR controller catches it. Tune Q and R and the Riccati equation is re-solved in real time.

Cart-pole · thesis A,BLQR LIVE

Controls

Drag the pendulum to disturb it. The gain K is recomputed from the algebraic Riccati equation whenever you move a slider.

angle θ0.0°
cart x0.00
force u0.0
statusbalanced
Turn the controller off and the pole falls — the open-loop system has an eigenvalue at +4.43 (right half-plane). LQR pulls all four poles into the left half-plane.
04
04 Rig II · One Aircraft

Single rigid body — the roll-yaw instability

A single aircraft with seven inputs (aileron, elevator, rudder, thrust + three gusts) and nine outputs. Left alone, its lateral motion diverges — roll and yaw wander off exactly as in the dissertation's Figure 12. Switch the LQR controller on and the roll snaps back to level.

Lateral roll-yaw · behind-viewLIVE

Controls

The horizon shows the aircraft's roll from behind; the trace plots roll angle over time.

roll φ0.0°
yaw ψ0.0°
unstable λ+0.00
modeDIVERGING
uncontrolled LQR controlled
05
05 Rig III · Coupled Flight

Twin body — coupling, gusts & structural LQR

Two aircraft joined at a hinge. Now the bodies fight each other: a disturbance on one rolls the other. Standard LQR assumes clean decoupling; structural LQR respects the real cross-coupling between control inputs and outputs — and holds both bodies steady even through a gust.

Two bodies · hinge couplingLIVE

Controls

Watch both bodies. A gust hits body 1 — see how it bleeds into body 2 through the hinge, and how control contains it.

roll 10.0°
roll 20.0°
controllernone
peak roll0.0°
body 1 body 2 gust
Cycle the control button: NONE → LQR → STRUCTURAL LQR. Structural LQR reacts more with the ailerons that actually drive roll, settling faster.
06
06 Rig IV · Variable Geometry

A-configuration vs V-configuration

The hinge angle reshapes the aircraft. A raised hinge (+17°) gives the A-configuration; a lowered hinge (−17°) gives the V-configuration — a genuinely non-linear change in dynamics. Sweep the hinge and compare how no control, LQR, and LQG handle each geometry.

Hinge geometry · roll responseLIVE

Controls

Drag the hinge slider between the two configurations. The plot shows roll response under three strategies; the geometry above morphs to match.

configA-configuration
hinge+17°
open-loop λ+0.11
stabilisedLQR+LQG
no control LQR LQG
The V-configuration is the more erratic of the two — it needs more aileron authority, exactly as reported. LQG's Kalman estimate smooths the response versus raw LQR.
07
07 Frequency Domain

Bode & Nyquist — the system's fingerprint

Time plots show what happens; frequency plots show why. The Bode curves reveal how strongly each input frequency is amplified and how far it is delayed; the Nyquist trace maps the same response onto the complex plane. Sweep the frequency marker and watch the point travel both plots at once.

Bode · magnitude & phaseSWEEPING

Controls

Drag the frequency marker; the same ω is highlighted on the Bode curves and the Nyquist locus. Toggle the controller to see how feedback reshapes the response.

magnitude0.0 dB
phase0.0°
gain margin
peak |G|
A Nyquist locus that encircles the −1 point signals instability. Feedback pulls the curve away from it — the same story the eigenvalues tell, seen from another angle.
08
08 Optimal Estimation

The Kalman filter — seeing through noise

LQR assumes you can measure every state perfectly. Real sensors are noisy. The Kalman filter fuses a model prediction with noisy measurements to estimate the true state — and pairing it with LQR gives LQG. Crank the sensor noise and watch the estimate hold its nerve while the raw measurement scatters.

Truth vs measurement vs estimateESTIMATING

Controls

Grey dots are what the sensor sees. The indigo line is the filter's estimate; the white line is the truth it cannot see directly.

true roll0.0°
estimate0.0°
RMS error0.00°
raw noise σ0.30
truth measurement estimate
Turn the filter off and the controller acts on raw noise — the response jitters. This is exactly why the dissertation's LQG design outperforms raw LQR under realistic sensing.
09
09 The Stability Scope

Watch the poles move

Stability lives in the complex plane. Everything left of the vertical axis is stable; anything to the right diverges. Pick a system and drag the control weight — the eigenvalues are recomputed live and slide across the boundary as the controller does its work.

Complex plane · s-domainλ LIVE

Controls

Green poles are stable, red poles are unstable. Raise the control weight to pull the poles into the left half-plane.

System stable — all eigenvalues in the left half-plane.
10
10 The Field

Literature map

The dissertation sits on a web of prior work — wingtip docking, formation flight, aerodynamic interaction, and modern control. Hover a node to trace its links; click to read the reference. Grouped by theme.

Hover a paper to see its role, or click a node for the full citation. The central node is this dissertation.
11
11 Outcomes

Findings & the code behind them

A progression from an unstable pole to a stabilised variable-geometry twin — each stage stabilised, analysed, and simulated. The conclusions, chapter by chapter:

The MATLAB behind the pendulum

The exact appendix routine — build A and B, solve the LQR, verify the closed-loop eigenvalues. This is the code the rig above reproduces in JavaScript.

inverted_pendulum_lqr.m

    
check_aircraft_stability.m