Lyapunov stability for roboticists

Linear systems have a stability test: check the eigenvalues of A. Done. Nonlinear systems don't have a one-shot test — you can have stable origins, regions of attraction, limit cycles, chaos. Lyapunov's idea (1892, still the standard 130 years later): pick an "energy-like" function of the state, show it always decreases, and you've proved the system converges. Once the trick clicks, every "is this controller stable?" paper becomes legible.

The intuition: energy

A pendulum with friction always stops at the bottom because its mechanical energy strictly decreases. Total energy $V$ starts positive, reaches zero at the rest state, and $\dot{V}\le 0$ along every trajectory. That's all you need.

Lyapunov generalized: any positive function with strictly negative derivative implies convergence. The function doesn't have to be physical energy — it just has to look like one.

The formal definition

For a system $$\dot{\mathbf{x}}=f(\mathbf{x})$$ with equilibrium at the origin, a function $V(\mathbf{x})$ is a Lyapunov function if:

1. $$V(0)=0$$
2. $$V(\mathbf{x})>0\text{for all }\mathbf{x}\ne 0$$
3. $$\dot{V}(\mathbf{x})=\nabla V\cdot f(\mathbf{x})\le 0\text{for all }\mathbf{x}$$

If V exists, the origin is stable (trajectories don't blow up). If $\dot{V}$ is strictly negative away from the origin, asymptotically stable (trajectories converge to zero). If both hold globally, globally asymptotically stable.

The simplest worked example

System: $$\dot{x}=-x^3$$. Try $$V(x)=\frac{1}{2}{x}^2$$.

Check: $V(0)=0$, $V(x)>0$ otherwise. Compute the derivative:

$$\dot{V}=x\dot{x}=-x^4\le 0$$

Strictly negative away from x = 0. The origin is globally asymptotically stable. Done. We never had to solve the differential equation.

Why this matters in practice

You'll use Lyapunov to:

• Prove a controller stabilizes the system. Standard pattern: define $V$ to be a quadratic in the error, design the control law so $\dot{V}\le 0$.
• Bound regions of attraction. The controller might only work near the equilibrium; the level sets of V tell you "trajectories starting inside this region converge."
• Design adaptive controllers. When system parameters are unknown, Lyapunov-based adaptation laws guarantee both tracking convergence and parameter consistency.
• Justify nonlinear-control choices. Sliding-mode, backstepping, energy shaping — all are Lyapunov-based.

The robot-arm example everyone meets first

Trajectory tracking on an arm with ideal dynamics. Define error $\mathbf{e}={\mathbf{\theta }}_d-\mathbf{\theta }$. Pick a Lyapunov candidate:

$$V(\mathbf{e},\dot{\mathbf{e}})=\frac{1}{2}{\dot{\mathbf{e}}}^{T}M(\mathbf{\theta })\dot{\mathbf{e}}+\frac{1}{2}{\mathbf{e}}^T{K}_p\mathbf{e}$$

That's "kinetic energy of the error" plus "elastic energy." Take the derivative, plug in the dynamics, choose torques τ that make $\dot{V}$ negative-definite. The result is a PD-plus-feedforward law that's provably stable for any positive-definite $K_p,K_d$.

You didn't tune by trial-and-error. You proved it from first principles. Slotine and Li's Applied Nonlinear Control works through this in detail — still the canonical reference.

The catch: finding V

Lyapunov gives sufficient conditions, not a recipe. You have to guess a candidate $V$. Some patterns:

• Quadratic forms $V={\mathbf{x}}^{T}P\mathbf{x}$ for linear systems. Solve the Lyapunov equation $A^{T}P+PA=-Q$ for any positive-definite Q. P exists iff A is stable.
• Energy-like functions for mechanical systems. Total energy or modified energy.
• Sum of squares (SOS). Modern computational technique: search over polynomial Lyapunov functions via convex optimization. Used by control researchers for systems where intuition fails.
• Composite Lyapunov: for cascaded or hierarchical systems, sum smaller V's together. Common in adaptive control.

For most robotics work, the first two suffice.

LaSalle's invariance principle (the trick when $$\dot{V}\le 0$$ but not strict)

Sometimes $\dot{V}=0$ on a non-trivial set (not just the origin). The trajectory could in principle hang there. LaSalle's principle says: if the only invariant subset of {x : $\dot{V}=0$} is the origin, you still have asymptotic stability.

This is what saves the arm-tracking proof above: $\dot{V}=-{\dot{\mathbf{e}}}^T{K}_d\dot{\mathbf{e}}$ only vanishes when $\dot{\mathbf{e}}=0$, but inspection of the dynamics shows the only invariant set is the origin. LaSalle gets you home.

What Lyapunov doesn't give you

• Global stability — only as far as $V$ is finite and $\dot{V}\le 0$.
• Exact convergence rate — Lyapunov-based exponential bounds exist but are usually loose.
• Robustness to disturbances — that's "input-to-state stability" (ISS), Lyapunov's modern relative.
• The actual controller — finding $V$ doesn't tell you τ; you still have to design the law so $\dot{V}\le 0$.

Lyapunov in 2026 robotics

• Stability proofs in safe RL. Recent work (Berkenkamp, Khansari-Zadeh) constrains learned policies to remain inside Lyapunov-stable sets.
• Control barrier functions. Modern descendant of Lyapunov, but for safety constraints rather than convergence. $V$ is replaced by $h(\mathbf{x})\ge 0$ with the constraint that $\dot{h}$ respects the boundary.
• Bipedal walking. Provably stable walking controllers on Cassie, Digit, Atlas use Lyapunov-based gaits.
• Adaptive control of soft robots. Heavy nonlinearity, parameter uncertainty — Lyapunov-based adaptation handles both.

What to learn first

The minimum you need to read papers:

1. The three conditions (positive function, zero at equilibrium, derivative non-positive).
2. Quadratic Lyapunov functions for linear systems via the Lyapunov equation.
3. Energy-based Lyapunov for mechanical systems (the arm example above).
4. LaSalle's principle for the "weak $\dot{V}$" case.

Slotine and Li's textbook (free PDFs floating around) covers all four in 50 pages. After that you can read most modern nonlinear-control papers.

Exercise

For a 1-DOF pendulum with friction ($$\ddot{\theta }+b\dot{\theta }+g\sin \theta /L=0$$), construct an energy-based Lyapunov function. Show that without friction it's marginally stable ($$\dot{V}=0$$); with friction it's asymptotically stable. Then add a torque-based stabilizing controller and re-derive — you should find an explicit gain condition for global asymptotic stability.

Next

Underactuated swing-up — when even Lyapunov isn't enough because there are fewer actuators than degrees of freedom.