Slung payload · rotor failure
Supplementary simulation study

Quasi-Static Fault-Tolerant Feedback Control of a Quadrotor Carrying a Slung Payload under Rotor Failure

A quadrotor carries a payload on a massless rod pivoted at its center of mass. One rotor has failed completely, so three propellers remain and the roll torque is tied to the thrust and the yaw torque. The quasi-static feedback of the paper is designed for the faulty dynamics. It drives the payload onto a geometric path, keeps the path invariant, assigns the motion along it and keeps the yaw rate bounded. This page carries a Drake simulation of the scenario of the paper, the figures of the paper regenerated from it, a run that starts on the path, a comparison of three implementations of the closed loop, and everything needed to reproduce them.

Scenario

The scenario of the paper

The path is the circle \(h_1 = p_1^2 + p_2^2 - d^{*2} = 0\), \(h_2 = p_3 - r^* = 0\) with \(d^* = 5\) m and \(r^* = -3\) m in the north-east-down frame. The path coordinate is the polar angle \(h_3 = \mathrm{atan2}(p_2, p_1)\) and its rate is driven to \(v_d = 0.8\) rad/s, so the payload moves at 4 m/s along the circle. The payload starts at \(p(0) = (-2, 8, -11)\) m with the pendulum at \(\alpha = -30^\circ\), \(\beta = 20^\circ\) and the quadrotor at \(\theta = 45^\circ\), \(\phi = 20^\circ\), \(\psi = -60^\circ\), with the velocities of Section VI. The controller is the quasi-static feedback (17), (26) of the paper with the damping (2), and the gains place the poles of the three chains at the values of the paper.

Snapshots of the Drake run at t = 0, 1, 1.5, 2, 5, 6.5, 8 and 10.5 s, the times of Fig. 2 of the paper, composed from one camera. The white line is the trace of the payload up to t = 10.5 s.
Fig. 2 of the paper from the Drake log. The same times, drawn in the north-east-down axes of the paper. The quadrotor is drawn larger than life, the red ring is the failed rotor and the black dot is the payload.
Outputs and inputs

Outputs and inputs

The two transverse outputs decay to zero and the speed error vanishes, as the transverse dynamics (31) of the paper require. The thrust settles at 52.8 N, a little above the weight because the pendulum leans inward on the circle. After the transient the torques stay small. The roll torque \(\tau_1 = c_1 u_t - c_2\tau_3\) is not free: it is set by the thrust and the yaw torque through the failure.

Fig. 4 of the paper from the Drake log. The outputs \(h_1\), \(h_2\), the unwrapped path coordinate \(h_3\) and the speed error \(\dot h_3 - v_d\). The payload is within 10 cm of the path after 6.9 s, within 1 cm after 9.2 s and within 1 mm after 11.4 s. The speed error is below 1 % of \(v_d\) after 7.3 s and below 5.1×10-4 rad/s after 10 s.
Fig. 3 of the paper from the Drake log. The thrust and the three applied torques, with the torque axes clipped to ±0.05 Nm as in the paper. The thrust ranges over [12.7, 62.6] N and the torques peak at (15.7, 20.1, 1.5) Nm in the first second. After 5 s they stay below (0.038, 0.037, 0.004) Nm.
Configuration

Configuration, pendulum and attitude

The payload position, the pendulum angles and the Euler angles of the quadrotor, then the pendulum direction with its angular velocity and the nine entries of the rotation matrix. The yaw angle grows without bound, by 70 rad over the run, while its rate stays bounded. This is the internal dynamics of Theorem 1: the yaw rate and the state \(\xi_2 = J_1\Omega_1 + J_3 c_2\Omega_3\) are bounded along the solution.

Fig. 5 of the paper from the Drake log. Payload position, pendulum angles and 2-1-3 Euler angles. The pitch and roll oscillate at the period of the circle once the payload is on it.
Fig. 6 of the paper from the Drake log. The pendulum direction \(q\) and its angular velocity \(\omega\). The pendulum never swings further than 31.5 degrees from the vertical.
Fig. 7 of the paper from the Drake log. The rotation matrix of the quadrotor. The first two columns turn at the yaw rate. The third column is \(b_3\), which does not depend on the yaw: it carries the small tilt that holds the payload on the circle and turns with the period of the circle, and \(R_{33}\) stays near 0.97.
The internal state. Left: the yaw rate \(\dot\psi\) and the body rate \(\Omega_3\), which settle at 3.55 rad/s with a peak of 4.26 rad/s. Middle: the internal state \(\xi_2\), bounded by 2.02 Nms. Right: the two functions that must stay positive on \(X_d\). Over the run \(\langle b_3, q\rangle \ge 0.61\) and \(\cos\alpha\cos\beta \ge 0.79\), so the solution stays in the domain of the controller and \(\tilde A\) stays invertible, with \(|\det\tilde A| \ge 83,048\).
Invariance

Starting on the path

At hover with the payload on the path, the pendulum vertical and the quadrotor level, every transverse coordinate is zero, so the state lies on the path-following manifold whatever the heading. The paper states that from such a start the transient keeps the payload on the path. In this run the payload starts at rest on the circle and the tangential input accelerates it to \(v_d\). The transverse coordinates stay at zero in continuous time, and in the sampled Drake run the payload never leaves the circle by more than 0.23 mm.

Distance of the payload to the path and the speed error for the start of the paper and for the hover start on the path, on a logarithmic scale. The sharp dips are the instants at which a signed error passes through zero. From the hover start the distance stays below 0.23 mm for the whole run, which is the residual set by the rotational inertia of the spherical payload, while the speed error decays from 0.8 rad/s and is below 1 % of \(v_d\) after 7.3 s.
0.23 mm
largest distance of the payload to the path over the run from the hover start
[50.9, 52.8] N
thrust range from the hover start, against 51.0 N of weight
0.50 Nm
largest applied torque from the hover start, against 20.1 Nm from the start of the paper
3.59 rad/s
largest yaw rate from the hover start
Implementations

Three implementations of the closed loop

The MATLAB code of the paper integrates the coordinate-free model (1) with ode45 and evaluates the controller from expressions generated by Maple. The Python port evaluates the same expressions, translated line by line, on the same model. At the initial condition and at 59 random states its thrust, torques, output chains and vector field agree with MATLAB to a relative 7.8×10-12, and the two closed-loop runs with tight tolerances agree to 4.8×10-10 m in the payload position. A verification gate checks the expressions against the closed loop itself: the time derivative of every term of the two sixth-order chains equals the next term, and at the top of each chain it equals the commanded \(\nu_1\) or \(\nu_3\), which tests the drift terms and the matrix \(\tilde A\) together. The Drake run replaces the model by a multibody plant: a free rigid body for the quadrotor and one rigid body for the pendulum, joined by a universal joint at the center of mass of the quadrotor, with the controller sampled at 1000 Hz. A second gate checks that the accelerations of this plant equal those of the model (1) at random states and inputs to 2.6×10-14 relative, with a point-mass payload.

Drake against the port. Left: the payload position difference, which is an offset along the circle. Middle: the thrust difference. Right: the distance to the path of both runs, with a dip where the Drake run crosses the path before settling on its residual. The offset along the circle comes from the zero-order hold: the body spins at 3.6 rad/s while the command is held, which biases the speed along the path by 5.3×10-4 rad/s at 1 kHz and 0.0012 rad/s at 400 Hz, measured as the mean difference of \(\dot h_3\) after 10 s. The residual distance to the path of 0.23 mm comes from the payload being a sphere: with a point mass the same run settles to 8.3×10-4 mm.
Port, point massDrake, 1 kHzDrake, point massDrake, 400 Hz
Payload within 1 cm of the path after [s]9.289.259.289.26
Distance to the path after 10 s [mm]< 4.9< 4.7< 5.0< 4.8
Distance to the path at 20 s [mm]2.3×10-40.238.3×10-40.23
Speed error after 10 s [rad/s]< 6.4×10-4< 5.1×10-4< 4.7×10-4< 0.0012
Yaw rate after 10 s [rad/s]3.553.553.553.55
Thrust after 5 s [N][52.81, 53.92][52.81, 53.91][52.81, 53.91][52.80, 53.89]
Against the port
Transverse difference after 10 s [mm]0.270.050.25
Offset along the circle at 20 s [mm]6662159
Thrust difference after 5 s [N]0.0090.0090.022
Torque difference after 5 s [Nm]2.1×10-42.1×10-45.4×10-4
Mean speed difference after 10 s [rad/s]-5.3×10-4-4.8×10-4-0.0012

The port integrates the coordinate-free model with a tolerance of 10-9. The Drake plant is integrated with a third-order Runge-Kutta scheme at 4 kHz. The spherical payload adds a rotational inertia of 2.9×10-4 kg m2 to the 1.8 kg m2 of the point mass about the pivot, and the universal joint lets it swing but not spin about the rod.

Verification gatesValue
Port against MATLAB, both with the default tolerances of ode45: largest payload position difference26 mm
Port against MATLAB, both with tight tolerances: largest payload position difference0.000 mm
Port against MATLAB, tight tolerances: largest thrust difference1.3×10-8 N
Controller, torques and vector field at 60 states against MATLABrelative difference 7.8×10-12
Chain gate, largest relative mismatch of a derivative along the closed loop6.9×10-10
Plant gate, accelerations of the Drake plant against the model (1)relative difference 2.6×10-14
Parameters

Every parameter

These are the complete settings behind every figure on this page. They are those of Section VI of the paper.

Slung payload system

ParameterValue
Quadrotor mass and inertiamq = 5 kg, J = diag(0.03, 0.03, 0.05) kg m2
Payload and rodmp = 0.2 kg at L = 3 m from the pivot on a massless rod. In Drake the payload is a sphere of radius 0.06 m
Rotor distance and drag ratioℓ = 0.2 m, cd = 0.01, so c1 = 0.1 m and c2 = 10
Failurethe rotor on the +b2 arm is lost, τ1 = c1 ut − c2 τ3
DampingkΩ = 3 s-1 on every axis, applied torque τa = τ − kΩ J Ω
Gravity9.81 m/s2, north-east-down frame

Path and controller

ParameterValue
Pathx2 + y2 = 52, z = -3 m, path coordinate h3 = atan2(p2, p1)
Speed referencevd = 0.8 rad/s (4 m/s along the circle)
Poles of the h1 chain (order 6){-1.0, -1.5, -2.0, -2.5, -3.0, -3.5}, K1 = (78.75, 250.875, 319, 208.125, 73.75, 13.5)
Poles of the h2 chain (order 2){-1.0, -1.2}, K2 = (1.2, 2.2)
Poles of the speed chain of h3 (order 5){-1.0, -1.5, -2.0, -2.5, -3.0}, K3 = (22.5, 65.25, 72.5, 38.75, 10)
Feedbackthrust from (17), torques τ2, τ3 from (26), τ1 from the failure, then the damping

Start of the paper

ParameterValue
Payloadp(0) = (-2, 8, -11) m, v(0) = (-0.2, 0.1, 0.5) m/s
Pendulumα = -30°, β = 20°, rates (-0.2, 0.1) rad/s
Quadrotorθ = 45°, φ = 20°, ψ = -60°, Ω(0) = (-0.1, 0.3, 0.2) rad/s
Horizon20 s

Hover start on the path

ParameterValue
Statepayload at rest on the circle, pendulum vertical, quadrotor level with zero rates. Every transverse coordinate is zero

Simulation

ParameterValue
MATLABcoordinate-free model (1), ode45 with its default tolerances, as in the paper
Python portthe same model and expressions, Dormand-Prince 8(5,3) with relative tolerance 10-9
DrakeMultibodyPlant, continuous, third-order Runge-Kutta at 4 kHz, controller at 1000 Hz with a zero-order hold on the body-frame thrust and torque
RenderingVTK render engine of Drake. The airframe is the Skydio X2 mesh of the MuJoCo Menagerie (Apache-2.0), scaled so that the rotor hubs sit at ℓ, and the airframe, rod and payload are drawn at three times their size
Reproduce

Reproducing this

The simulation code is in the repository under sim/ and the MATLAB code of the paper under matlab/. The verification gates run first.

# set up and verify
git clone https://github.com/gradslab/qs-sl.git
cd qs-sl/sim
pip install -r requirements.txt
python3 model/test_chains.py
python3 model/compare_matlab.py
python3 drake/verify_plant.py

# the runs behind every number and figure
python3 model/simulate_port.py --tight
python3 drake/sim_drake.py
python3 drake/sim_drake.py --start on-path --tag _onpath
python3 drake/sim_drake.py --point-mass --tag _pointmass
python3 drake/sim_drake.py --hz 400 --tag _400hz

# the numbers and the figures
python3 metrics.py
python3 paper_figures.py
python3 paper_figures.py --web

# animations (Drake VTK renderer and ffmpeg), then the page
python3 drake/render_drake.py drake_run --snapshots 0,1,1.5,2,5,6.5,8,10.5 --cam wide
for c in follow side orbit; do python3 drake/render_drake.py drake_run --cam $c; done
python3 drake/render_drake.py drake_run_onpath --cam follow
python3 web/anim_web.py
python3 web/page_numbers.py
python3 web/build_page.py
Citation

Citation

@unpublished{allawati2026quasistatic,
  author = {Mohamed Al Lawati and Adeel Akhtar},
  title  = {Quasi-Static Fault-Tolerant Feedback Control of a Quadrotor
            Carrying a Slung Payload under Rotor Failure},
  note   = {Under review},
  year   = {2026}
}