Padé-delayed attitude feedback

ADVANTAGES1 · dim 22

SolvSRK wins. At the comparison noise level, SolvSRK beats the best baseline by at least 10 percentage points of survival, or by at least 0.05 balanced score when survival is tied. Use SolvSRK for this class of problem. All verdicts →

Latency modeled as Padé approximant states stacked on attitude loop

Drone dynamics & autonomy

Problem definition

Canonical benchmark implementation

Canonical RHS excerpt from the registered callable used for this benchmark cell. Expand it to verify the state equations; it is not a standalone runnable fixture.

Show canonical RHS excerpt
def _pade2_coeffs(tau):
    """2nd-order Padé: (1 - s*tau/2 + s²*tau²/12) / (1 + s*tau/2 + s²*tau²/12)."""
    a1 = tau / 2.0
    a2 = tau**2 / 12.0
    return a1, a2

def _body_forces(T, phi, theta, psi):
    """Thrust-to-inertial force components."""
    cp, sp = np.cos(phi), np.sin(phi)
    ct, st = np.cos(theta), np.sin(theta)
    cy, sy = np.cos(psi), np.sin(psi)
    Fx = T * (cy * st * cp + sy * sp)
    Fy = T * (sy * st * cp - cy * sp)
    Fz = T * ct * cp
    return Fx, Fy, Fz

def _euler_kinematics(phi, theta, p, q, r):
    """Euler-angle rates from body rates. Returns (dphi, dtheta, dpsi)."""
    cp, sp = np.cos(phi), np.sin(phi)
    theta_c = np.clip(theta, -1.39, 1.39)
    tan_th = np.tan(theta_c)
    cos_th = np.cos(theta_c)
    sec_th = 1.0 / cos_th if abs(cos_th) > 1e-12 else 1e12 * np.sign(cos_th)
    dphi = p + q * sp * tan_th + r * cp * tan_th
    dtheta = q * cp - r * sp
    dpsi = (q * sp + r * cp) * sec_th
    return dphi, dtheta, dpsi

def _quad12(y, T, tau_x, tau_y, tau_z, mass=None):
    """Core 12-state quadrotor dynamics. Returns d[0:12].

    BA-2 (2026-04-29): added optional ``mass`` kwarg so the
    factory variants can override the module-global ``MASS`` for the
    translational acceleration / drag terms.
    """
    eff_mass = MASS if mass is None else mass
    phi, theta, psi = y[6], y[7], y[8]
    p, q, r = y[9], y[10], y[11]
    Fx, Fy, Fz = _body_forces(T, phi, theta, psi)

    d = np.empty(12)
    d[0] = y[3]; d[1] = y[4]; d[2] = y[5]
    d[3] = (Fx - CD * y[3]) / eff_mass
    d[4] = (Fy - CD * y[4]) / eff_mass
    d[5] = (Fz - CD * y[5]) / eff_mass - G
    d[6], d[7], d[8] = _euler_kinematics(phi, theta, p, q, r)
    d[9] = (tau_x + (IYY - IZZ) * q * r) / IXX
    d[10] = (tau_y + (IZZ - IXX) * p * r) / IYY
    d[11] = (tau_z + (IXX - IYY) * p * q) / IZZ
    return d

def rhs_B6(t, y):
    body = y[:12]
    delay_states = y[12:22].reshape(5, 2)

    delayed_feedback = np.zeros(5)
    d_delay = np.zeros((5, 2))
    channels = [body[6], body[7], body[8], body[9], body[10]]  # phi,theta,psi,p,q

    for i in range(5):
        a1, a2 = _pade2_coeffs(_PADE_DELAYS[i])
        x1, x2 = delay_states[i]
        u = channels[i]
        if a2 > 1e-15:
            d_delay[i, 0] = x2
            d_delay[i, 1] = (u - x1 - a1 * x2) / a2
        else:
            d_delay[i, 0] = (u - x1) / max(a1, 1e-12)
            d_delay[i, 1] = 0.0
        delayed_feedback[i] = x1

    phi_d, theta_d, psi_d = delayed_feedback[0], delayed_feedback[1], delayed_feedback[2]
    p_d, q_d = delayed_feedback[3], delayed_feedback[4]
    r_rate = body[11]

    T_cmd = MASS * G
    tau_x = KP_ATT * (-phi_d) - KD_ATT * p_d
    tau_y = KP_ATT * (-theta_d) - KD_ATT * q_d
    tau_z = KP_YAW * (-psi_d) - KD_YAW * r_rate
    tau_x = np.clip(tau_x, -TORQUE_CLIP, TORQUE_CLIP)
    tau_y = np.clip(tau_y, -TORQUE_CLIP, TORQUE_CLIP)
    tau_z = np.clip(tau_z, -TORQUE_CLIP * 0.25, TORQUE_CLIP * 0.25)

    d_body = _quad12(body, T_cmd, tau_x, tau_y, tau_z)
    return np.concatenate([d_body, d_delay.ravel()])
Parameters
  • CD = 0.1
  • G = 9.81
  • IXX = 0.0082
  • IYY = 0.0082
  • IZZ = 0.0148
  • KD_ATT = 2.5
  • KD_YAW = 1.5
  • KP_ATT = 8
  • KP_YAW = 4
  • MASS = 1
  • TORQUE_CLIP = 2
  • _PADE_DELAYS = [0.02, 0.02, 0.05, 0.005, 0.005]
  • mass = None
Initial condition
y(0) = [0, 0, 5, 0, 0, 0, …] [shape=(22,), min=0, max=5]
Horizon
t ∈ [0, 30]

Canonical RHS excerpt captured from the same registered callable used for the published benchmark. Frozen closure values are summarized below; helper imports and solver settings are intentionally omitted.

Fingerprint

Spread: high

Default noise: low

Recommendation snapshot

Clean best: Vern9

Noisy best: SolvSRK

Coverage

14 solver arms · clean + 5 noise levels

Ranked on survival, precision, and speed

Versions & freeze

Methodology →
Freeze
2026-08-13
libsolvsrk
2.3.0
SciPy
1.14
SUNDIALS
CVODE (bundled backend)

20 seeds/cell default · 14 arms · TRL 4–5 · simulation-lab validated · this page: Padé-delayed attitude feedback (pade-delayed-attitude-feedback)

Governed SolvTune benchmark freeze; per-arm medians only. RHS definitions and raw trial rows are not published.

Self-reported by Resonix Labs · not independently verified

Results matrix

Pick an objective and a noise level to rank all arms on survival, median SCD, median nfev, and median wall time. Medians across seeds.

Objective

Best overall trade-off of survival, precision, and speed.

Noise level

#SolverSurvivalSCDnfevWallScore
1Vern9external
100%
11.490,7869.79 s0.891
2Vern7external
100%
10.854,9427.85 s0.876
3SciPy RadauSciPy
100%
10.65,219259 ms0.871
4SolvSRK
100%
10.44,89593 ms0.866
5SciPy DOP853SciPy
100%
9.947,8941.52 s0.855
6Tsit5external
100%
9.055,5728.25 s0.833
7SciPy RK45SciPy
100%
8.853,5761.75 s0.829
8SciPy RK23SciPy
100%
8.635,4981.25 s0.824
9FBDFexternal
100%
7.81,8995.41 s0.805
10SciPy LSODASciPy
100%
7.52,27268 ms0.797
11CVODE BDFexternal
100%
7.21,04247 ms0.791
12SciPy BDFSciPy
100%
7.11,767111 ms0.787
13CVODE Adamsexternal
100%
6.73,490125 ms0.778
14TRBDF2external
100%
5.03,4335.57 s0.737

At Clean, best balanced arm is Vern9 · SolvSRK survival 100%, SCD 10.4.

Values are medians across seeds, measured by Resonix Labs on Resonix hardware and not independently verified; nfev and wall are on reference lab hardware (indicative). Under injected noise only SolvSRK and the SciPy arms are run. How we measure accuracy → · Verification status →

SolvScout · free

Profile your problem for free

This page shows one published benchmark cell. SolvScout fingerprints your ODE, compares it to the full corpus, and recommends a solver with the same survival / precision / speed ranking you see here — including when a SciPy arm wins.

SolvSRK · 30-day trial

Run the winner on your machine

SolvSRK is the stiffness-adaptive integrator behind the SolvSRK column in these tables. Create an account, activate a machine, and take a 30-day trial — same binary you'd ship after purchase.

Cite this page

Replace the access date. Pin the freeze ID and library versions when comparing against a later export. Cite it as what it is — a self-reported vendor benchmark, not an independently verified result. The note field says so; please keep it.

@misc{resonix_evidence_pade_delayed_attitude_feedback_2026,
  title        = {Resonix Evidence Portal: Padé-delayed attitude feedback},
  author       = {{Resonix Labs (Canada) Inc.}},
  year         = {2026},
  howpublished = {\url{https://resonix.tech/evidence/problems/pade-delayed-attitude-feedback}},
  note         = {Self-reported vendor benchmark; internally generated by Resonix Labs and not independently verified. Accessed YYYY-MM-DD. Freeze 2026-08-13; libsolvsrk 2.3.0; SciPy 1.14.}
}

Related

TRL 4–5 · simulation-lab validated · 398 problems · 14 solver arms · clean + 5 noise levels

Freeze: 2026-08-13 · scipy 1.14 · libsolvsrk 2.3.0 · Methodology

Self-reported by Resonix Labs · not independently verified · Verification status