1. Why Nonlinear Attitude Control Is Its Own Topic

Spacecraft attitude control is nonlinear even before you add flexible appendages or real actuators. The reason is that attitude is not a Euclidean vector: it evolves on a rotation manifold, the rigid-body dynamics include nonlinear cross-coupling, and the actuators and digital implementation introduce discontinuities and constraints.

• Geometry is nonlinear: attitude lives on the rotation manifold; with quaternions you must satisfy $\lVert q \rVert = 1$ at all times.
• Dynamics are nonlinear: rigid-body rotation includes the gyroscopic coupling term $\boldsymbol{\omega}\times(\mathbf{J}\boldsymbol{\omega})$.
• Actuators are nonlinear: torque saturation, rate limits, deadbands, minimum impulse bit, and thruster pulsing.
• Real loops include delays + filters: onboard sampling and processing delay can shrink phase margin and invite oscillations.
• Flexibility + discontinuities create limit cycles: pulsed thrusters / saturations can inject energy near structural modes.

So a “linear PD around a point” can behave perfectly in a small-angle lab test and still misbehave for large-angle slews, saturation-dominated maneuvers, or sign changes in attitude error. Domain D.7 designs the controller directly on quaternion geometry and then makes it implementable.

2. Core Model Used Throughout D.7

2.1 Quaternion kinematics (attitude propagation)

Let spacecraft attitude be the unit quaternion

\[ q = \begin{bmatrix}\mathbf{q}_v \\ q_4 \end{bmatrix}, \qquad \lVert q \rVert = 1 \]

where $\mathbf{q}_v\in\mathbb{R}^3$ is the vector part and $q_4$ is the scalar part. With body angular rate $\boldsymbol{\omega}$,

\[ \dot{q} = \frac{1}{2}\,\Omega(\boldsymbol{\omega})\,q \]

with a standard structure

\[ \Omega(\boldsymbol{\omega})= \begin{bmatrix} -[\boldsymbol{\omega}]_\times & \boldsymbol{\omega} \\ -\boldsymbol{\omega}^\mathsf{T} & 0 \end{bmatrix} \]

Key message: kinematics are linear in $q$ but nonlinear overall because $\Omega$ depends on $\boldsymbol{\omega}$ and we must enforce $\lVert q\rVert = 1$.

2.2 Rigid-body rotational dynamics (attitude plant)

\[ \mathbf{J}\,\dot{\boldsymbol{\omega}} + \boldsymbol{\omega}\times(\mathbf{J}\boldsymbol{\omega}) = \boldsymbol{\tau} + \boldsymbol{\tau}_d \]

• $\mathbf{J}$: inertia matrix (symmetric, positive definite)
• $\boldsymbol{\tau}$: control torque (reaction wheels / CMGs / thrusters / magnetorquers)
• $\boldsymbol{\tau}_d$: disturbances (gravity-gradient, SRP, drag, magnetic residuals, modeling error)

The plant is nonlinear because of $\boldsymbol{\omega}\times(\mathbf{J}\boldsymbol{\omega})$. Even if $\boldsymbol{\tau}$ is “PD,” the closed-loop is still nonlinear.

3. Quaternion Tracking Error (the error signal that behaves well)

Given desired attitude $q_d$, define the error quaternion:

\[ q_e = q_d^{-1}\otimes q \;\;\Rightarrow\;\; q_e = \begin{bmatrix} \mathbf{q}_{ev}\\ q_{e4} \end{bmatrix} \]

Interpretation: $q_e$ represents the rotation that takes you from desired attitude to current attitude.

3.1 Why we use the vector part $\mathbf{q}_{ev}$

The vector part $\mathbf{q}_{ev}$ points along the instantaneous axis that “undoes” the attitude error. It is the most practical geometric error for feedback: it directly encodes the direction your control torque should push.

3.2 Shortest-rotation convention (avoid unwinding)

Because $q$ and $-q$ represent the same physical attitude, the error quaternion has the same ambiguity. If you do nothing, the controller can rotate the long way around (“unwinding”).

A standard operational fix:

\[ \text{if } q_{e4} < 0,\quad q_e \leftarrow -q_e \]

This makes the controller choose the shorter rotation consistently.

3.3 Rate error

For setpoint regulation, $\boldsymbol{\omega}_d=0 \Rightarrow \boldsymbol{\omega}_e=\boldsymbol{\omega}$. For tracking maneuvers, $\boldsymbol{\omega}_e$ compares measured rate to the desired rate mapped into the correct frame.

4. Quaternion Feedback Control Law (baseline nonlinear stabilizer)

4.1 The practical quaternion PD law

The workhorse for spacecraft attitude regulation is:

\[ \boldsymbol{\tau} = -\mathbf{K}_p\,\mathbf{q}_{ev} - \mathbf{K}_d\,\boldsymbol{\omega} \]

• $-\mathbf{K}_p\,\mathbf{q}_{ev}$: restoring torque in quaternion space (like a spring in rotation geometry)
• $-\mathbf{K}_d\,\boldsymbol{\omega}$: damping (removes kinetic energy, prevents ringing)

This is PD-like but not a small-angle linearization. With correct sign handling of $q_e$, it remains meaningful for large slews.

4.2 Adding a model-based term (optional but strong)

If you have good estimates of $\mathbf{J}$ and desired rate/acceleration profiles, you can add a computed-torque style feedforward:

\[ \boldsymbol{\tau} = \boldsymbol{\omega}\times(\mathbf{J}\boldsymbol{\omega}) + \mathbf{J}\dot{\boldsymbol{\omega}}_d - \mathbf{K}_p\,\mathbf{q}_{ev} - \mathbf{K}_d\,\boldsymbol{\omega}_e \]

• first two terms “pre-cancel” nonlinear coupling and follow commanded accelerations
• PD terms supply robustness when the model is imperfect and suppress noise/uncertainty

In flight software, this is often implemented partially (or with conservative $\mathbf{J}$) so you don’t amplify sensor noise.

5. Lyapunov Stability Proof Structure (how you justify nonlinear control)

A clean proof is not about showing pretty algebra; it is about showing the controller dissipates energy despite nonlinear dynamics.

5.1 Choose a Lyapunov candidate (energy-like)

A very common structure is:

\[ V(\boldsymbol{\omega}, q_e) = \frac{1}{2}\boldsymbol{\omega}^\mathsf{T}\mathbf{J}\boldsymbol{\omega} + V_q(q_e) \]

• $\frac{1}{2}\boldsymbol{\omega}^\mathsf{T}\mathbf{J}\boldsymbol{\omega}$ is rotational kinetic energy
• $V_q(q_e)$ is an attitude “potential” that is minimal at zero attitude error

A widely used potential is shaped so that: it is minimal at $q_{e4}=1$ (no error), it respects quaternion geometry, it avoids unwinding by using the sign convention $q_{e4}\ge 0$.

5.2 Differentiate along system trajectories

Using $\mathbf{J}\dot{\boldsymbol{\omega}} = -\boldsymbol{\omega}\times(\mathbf{J}\boldsymbol{\omega}) + \boldsymbol{\tau} + \boldsymbol{\tau}_d$, the derivative typically takes the form:

\[ \dot{V} = \boldsymbol{\omega}^\mathsf{T}\boldsymbol{\tau} + \boldsymbol{\omega}^\mathsf{T}\boldsymbol{\tau}_d - \boldsymbol{\omega}^\mathsf{T}\mathbf{K}_d\boldsymbol{\omega} \]

where the “mixed” attitude terms cancel by how $V_q$ is chosen and by the structure of $\boldsymbol{\tau}=-\mathbf{K}_p\mathbf{q}_{ev}-\mathbf{K}_d\boldsymbol{\omega}$.

5.3 What the inequality tells you

If $\boldsymbol{\tau}_d=0$:

\[ \dot{V} = -\boldsymbol{\omega}^\mathsf{T}\mathbf{K}_d\boldsymbol{\omega}\le 0 \]

so energy decreases and the motion settles.

If $\boldsymbol{\tau}_d\neq 0$ but bounded: you get practical stability — the system converges to a small neighborhood whose size depends on disturbance magnitude and gains. Increasing damping improves rate rejection.

5.4 What kind of stability you should claim

Because attitude is not a Euclidean space, smooth global stabilization has topological caveats. In practice you state:

• almost-global asymptotic stability (all initial attitudes except a measure-zero set), or
• global practical stability under bounded disturbances and actuator constraints.

That’s also why sign-handling $q_{e4}\ge 0$ is not a cosmetic trick — it directly affects global behavior and prevents long-path rotations.

6. PD Control in Quaternion Space (design and tuning mindset)

6.1 Small-angle intuition still helps (but you design for large angles)

For small errors, $\mathbf{q}_{ev}\approx \frac{1}{2}\theta\,\hat{\mathbf{u}}$. Then:

\[ \boldsymbol{\tau}\approx -\mathbf{K}_p\left(\frac{1}{2}\theta\,\hat{\mathbf{u}}\right) - \mathbf{K}_d\boldsymbol{\omega} \]

So:

• $\mathbf{K}_p$ sets how “stiff” the attitude correction feels
• $\mathbf{K}_d$ sets damping and suppresses overshoot and oscillation

But the tuning must be validated in the large-angle + saturation regime, because that is the true nonlinear operating condition.

6.2 A realistic tuning workflow

1. Choose $\mathbf{K}_d$ high enough to tame rate transients and avoid ringing under saturation.
2. Increase $\mathbf{K}_p$ to hit maneuver-time requirements (slew speed).
3. Re-check $\mathbf{K}_d$ for damping under worst-case inertia and maximum disturbance torque.

Verify for:

• large-angle slews (worst-case initial error)
• actuator saturation duration
• sensor noise (gyro noise + bias)
• delays (sample time, filtering)
• flexible-mode sensitivity (if appendages exist)

A controller that “feels perfect” in the linear regime can excite modes or create limit cycles once clipping starts.

7. Sliding Mode Control (intro: robustness for uncertainty and disturbances)

Sliding mode is used when you expect: uncertain inertia / changing mass properties, bounded but unknown disturbance torques, coarse switching actuators (thrusters), or unmodeled dynamics that degrade PD performance.

7.1 Define a sliding surface

A standard surface blends rate and attitude error:

\[ \mathbf{s} = \boldsymbol{\omega} + \mathbf{\Lambda}\,\mathbf{q}_{ev} \]

Goal: drive $\mathbf{s}\to 0$. Once on the surface, rates become “tied” to attitude error in a stable way, giving strong robustness.

7.2 Control structure (concept)

A typical form is:

\[ \boldsymbol{\tau} = \boldsymbol{\tau}_{eq} - \mathbf{K}_s\,\mathrm{sat}\!\left(\frac{\mathbf{s}}{\phi}\right) \]

• $\boldsymbol{\tau}_{eq}$: nominal model-based part
• $\mathbf{K}_s$: robustness gain (chosen to dominate bounded uncertainties)
• $\mathrm{sat}(\mathbf{s}/\phi)$: boundary-layer saturation (practical replacement for $\mathrm{sgn}(\mathbf{s})$)

7.3 Why boundary layers are essential in spacecraft

Ideal sliding mode uses discontinuous switching. Real spacecraft have sampling, quantization, delays, and flexible modes — so discontinuous switching can cause chattering, which:

• excites flexible structures
• injects jitter into pointing
• increases actuator wear (thruster cycling / wheel torque chatter)

Hence practical sliding mode is almost always smoothed (boundary layer, hysteresis, rate limits, or higher-order variants).

8. Actuator Limits (the difference between “stable on paper” and “flies well”)

8.1 Torque saturation (hard limit)

All actuators have $\lVert\boldsymbol{\tau}\rVert \le \tau_{\max}$. The moment the controller commands beyond that, the loop changes character:

• effective gains drop
• damping may be insufficient
• overshoot and limit cycles can appear
• structural modes can be excited by clipped or pulsed commands

So saturation must be treated as part of the design, not an afterthought.

8.2 Slew-rate limits and eigenaxis-style maneuvers

Many missions impose maximum allowable angular rate (star tracker tracking, payload smear, structural loads). Under a rate cap, a physically realistic maneuver naturally becomes:

1. Acceleration phase: build up to allowable rate
2. Coast phase: rotate near constant maximum rate
3. Deceleration phase: reduce rate to zero at target

A saturation-aware quaternion controller can be designed so that these phases emerge naturally — especially important when you want predictable maneuver timing and bounded stress.

8.3 Deadbands and minimum impulse bit (thruster reality)

Thrusters are switching devices:

• small commands may be ignored (deadband)
• there is a minimum pulse width / impulse bit
• the effective torque becomes piecewise and discontinuous

This is a direct pathway to nonlinear oscillations (limit cycles), especially if the spacecraft has flexible modes. That is why practical thruster attitude control often uses modulator logic (e.g., pulse-width / pulse-frequency modulation) plus structural filtering.

8.4 Structured saturation: cascade saturation (why it appears in nonlinear design)

Instead of one harsh clip, cascade saturation limits the command gradually in layers. Conceptually, it shapes the control so that:

• directionality is preserved as much as possible
• stability is maintained during saturation-dominated maneuvers
• large-angle eigenaxis-like rotation is not destroyed by abrupt clipping

This is especially valuable when the controller must remain stable across wide operating regimes.

8.5 Control allocation and redistribution (multi-actuator systems)

With multiple wheels or thrusters, commanded body torque must be allocated through a distribution matrix. When one actuator saturates:

• the remaining actuators must pick up the demand
• the net body torque direction should stay aligned with control intent
• the controller should be “aware” of achievable torque to avoid hidden authority loss

A flight-ready design closes the loop on what torque was actually achieved, not just what was commanded.

9. Implementation Rules That Belong Inside D.7 (what you actually code)

9.1 Quaternion normalization

Numerical integration drifts. Always enforce:

\[ q \leftarrow \frac{q}{\lVert q\rVert} \]

9.2 Shortest-rotation sign rule

Always apply:

\[ q_{e4}\ge 0 \quad (\text{flip } q_e \text{ if needed}) \]

This prevents unwinding and makes behavior consistent.

9.3 Discrete-time reality

Controllers are sampled. You must verify:

• stability margin with sample time
• noise sensitivity (rate feedback amplifies gyro noise)
• delay + filtering effects

9.4 Robustness checks

Always test:

• inertia uncertainty (configuration changes)
• disturbance torque envelopes
• saturation and rate limits
• sensor bias/noise
• flexible mode injection (if present)

10. What D.7 Should Teach (final learning outcomes)

After this section, the reader should be able to:

• Build quaternion error correctly and enforce shortest-path behavior (avoid unwinding).
• Implement quaternion PD for large-angle attitude stabilization and understand why it is nonlinear even if it looks “PD.”
• Explain Lyapunov proof structure: pick $V$, compute $\dot{V}$, show energy dissipation and what changes with disturbances.
• Understand sliding mode at an engineering level: surface, robustness term, and why chattering mitigation is mandatory in spacecraft.
• Treat actuator limits as part of control design: saturation, rate caps, deadbands, pulsing, allocation — because these dominate real flight behavior.