Underactuated Robotics

Algorithms for Walking, Running, Swimming, Flying, and Manipulation

Russ Tedrake

© Russ Tedrake, 2022
Last modified .
How to cite these notes, use annotations, and give feedback.

Note: These are working notes used for a course being taught at MIT. They will be updated throughout the Spring 2022 semester. Lecture videos are available on YouTube.

Previous Chapter Table of contents Next Chapter

Lyapunov Analysis

Optimal control provides a powerful framework for formulating control problems using the language of optimization. But solving optimal control problems for nonlinear systems is hard! In many cases, we don't really care about finding the optimal controller, but would be satisfied with any controller that is guaranteed to accomplish the specified task. In many cases, we still formulate these problems using computational tools from optimization, and in this chapter we'll learn about tools that can provide guaranteed control solutions for systems that are beyond the complexity for which we can find the optimal feedback.

There are many excellent books on Lyapunov analysis; for instance Slotine90 is an excellent and very readable reference and Khalil01 can provide a rigorous treatment. In this chapter I will summarize (without proof) some of the key theorems from Lyapunov analysis, but then will also introduce a number of numerical algorithms... many of which are new enough that they have not yet appeared in any mainstream textbooks.

Lyapunov Functions

Let's start with our favorite simple example.

Stability of the Damped Pendulum

Recall that the equations of motion of the damped simple pendulum are given by \[ ml^2 \ddot{\theta} + mgl\sin\theta = -b\dot{\theta}, \] which I've written with the damping on the right-hand side to remind us that it is an external torque that we've modeled.

These equations represent a simple second-order differential equation; in chapter 2 we discussed at some length what was known about the solutions to this differential equation--in practice we do not have a closed-form solution for $\theta(t)$ as a function of the initial conditions. Since we couldn't provide a solution analytically, in chapter 2 we resorted to a graphical analysis, and confirmed the intuition that there are fixed points in the system (at $\theta = k\pi$ for every integer $k$) and that the fixed points at $\theta = 2\pi k$ are asymptotically stable with a large basin of attraction. The graphical analysis gave us this intuition, but can we actually prove this stability property? In a way that might also work for much more complicated systems?

One route forward was from looking at the total system energy (kinetic + potential), which we can write down: \[ E(\theta,\dot{\theta}) = \frac{1}{2} ml^2\dot{\theta}^2 - mgl \cos\theta. \] Recall that the contours of this energy function are the orbits of the undamped pendulum.

A natural route to proving the stability of the downward fixed points is by arguing that energy decreases for the damped pendulum (with $b>0$) and so the system will eventually come to rest at the minimum energy, $E = -mgl$, which happens at $\theta=2\pi k$. Let's make that argument slightly more precise.

Evaluating the time derivative of the energy reveals \[ \frac{d}{dt} E = - b\dot\theta^2 \le 0. \] This is sufficient to demonstrate that the energy will never increase, but it doesn't actually prove that the energy will converge to the minimum when $b>0$ because there are multiple states(not only the minimum) for which $\dot{E}=0$. To take the last step, we must observe that set of states with $\dot\theta=0$ is not an invariant set; that if the system is in, for instance $\theta=\frac{\pi}{4}, \dot\theta=0$ that it will not stay there, because $\ddot\theta \neq 0$. And once it leaves that state, energy will decrease once again. In fact, the fixed points are the only subset the set of states where $\dot{E}=0$ which do form an invariant set. Therefore we can conclude that as $t\rightarrow \infty$, the system will indeed come to rest at a fixed point (though it could be any fixed point with an energy less than or equal to the initial energy in the system, $E(0)$).

This is an important example. It demonstrated that we could use a relatively simple function -- the total system energy -- to describe something about the long-term dynamics of the pendulum even though the actual trajectories of the system are (analytically) very complex. It also demonstrated one of the subtleties of using an energy-like function that is non-increasing (instead of strictly decreasing) to prove asymptotic stability.

Lyapunov functions generalize this notion of an energy function to more general systems, which might not be stable in the sense of some mechanical energy. If I can find any positive function, call it $V(\bx)$, that gets smaller over time as the system evolves, then I can potentially use $V$ to make a statement about the long-term behavior of the system. $V$ is called a Lyapunov function.

Recall that we defined three separate notions for stability of a fixed-point of a nonlinear system: stability i.s.L., asymptotic stability, and exponential stability. We can use Lyapunov functions to demonstrate each of these, in turn.

Lyapunov's Direct Method

Given a system $\dot{\bx} = f(\bx)$, with $f$ continuous, and for some region ${\cal B}$ around the origin (specifically an open subset of $\mathbf{R}^n$ containing the origin), if I can produce a scalar, continuously-differentiable function $V(\bx)$, such that \begin{gather*} V(\bx) > 0, \forall \bx \in {\cal B} \setminus \{0\} \quad V(0) = 0, \text{ and} \\ \dot{V}(\bx) = \pd{V}{\bx} f(\bx) \le 0, \forall \bx \in {\cal B} \setminus \{0\} \quad \dot{V}(0) = 0, \end{gather*} then the origin $(\bx = 0)$ is stable in the sense of Lyapunov (i.s.L.). [Note: the notation $A \setminus B$ represents the set $A$ with the elements of $B$ removed.]

If, additionally, we have $$\dot{V}(\bx) = \pd{V}{\bx} f(\bx) < 0, \forall \bx \in {\cal B} \setminus \{0\},$$ then the origin is (locally) asymptotically stable. And if we have $$\dot{V}(\bx) = \pd{V}{\bx} f(\bx) \le -\alpha V(x), \forall \bx \in {\cal B} \setminus \{0\},$$ for some $\alpha>0$, then the origin is (locally) exponentially stable.

Note that for the sequel we will use the notation $V \succ 0$ to denote a positive-definite function, meaning that $V(0)=0$ and $V(\bx)>0$ for all $\bx\ne0$ (and also $V \succeq 0$ for positive semi-definite, $V \prec 0$ for negative-definite functions).

The intuition here is exactly the same as for the energy argument we made in the pendulum example: since $\dot{V}(x)$ is always zero or negative, the value of $V(x)$ will only get smaller (or stay the same) as time progresses. Inside the subset ${\cal B}$, for every $\epsilon$-ball, I can choose a $\delta$ such that $|x(0)|^2 < \delta \Rightarrow |x(t)|^2 < \epsilon, \forall t$ by choosing $\delta$ sufficiently small so that the sublevel set of $V(x)$ for the largest value that $V(x)$ takes in the $\delta$ ball is completely contained in the $\epsilon$ ball. Since the value of $V$ can only get smaller (or stay constant) in time, this gives stability i.s.L.. If $\dot{V}$ is strictly negative away from the origin, then it must eventually get to the origin (asymptotic stability). The exponential condition is implied by the fact that $\forall t>0, V(\bx(t)) \le V(\bx(0)) e^{-\alpha t}$.

Notice that the system analyzed above, $\dot{\bx}=f(\bx)$, did not have any control inputs. Therefore, Lyapunov analysis is used to study either the passive dynamics of a system or the dynamics of a closed-loop system (system + control in feedback). We will see generalizations of the Lyapunov functions to input-output systems later in the text.

Global Stability

The notion of a fixed point being stable i.s.L. is inherently a local notion of stability (defined with $\epsilon$- and $\delta$- balls around the origin), but the notions of asymptotic and exponential stability can be applied globally. The Lyapunov theorems work for this case, too, with only minor modification.

Lyapunov analysis for global stability

Given a system $\dot{\bx} = f(\bx)$, with $f$ continuous, if I can produce a scalar, continuously-differentiable function $V(\bx)$, such that \begin{gather*} V(\bx) \succ 0, \\ \dot{V}(\bx) = \pd{V}{\bx} f(\bx) \prec 0, \text{ and} \\ V(\bx) \rightarrow \infty \text{ whenever } ||x||\rightarrow \infty,\end{gather*} then the origin $(\bx = 0)$ is globally asymptotically stable (G.A.S.).

If additionally we have that $$\dot{V}(\bx) \preceq -\alpha V(\bx),$$ for some $\alpha>0$, then the origin is globally exponentially stable.

The new condition, on the behavior as $||\bx|| \rightarrow \infty$ is known as "radially unbounded", and is required to make sure that trajectories cannot diverge to infinity even as $V$ decreases; it is only required for global stability analysis.

LaSalle's Invariance Principle

Perhaps you noticed the disconnect between the statement above and the argument that we made for the stability of the pendulum. In the pendulum example, using the mechanical energy resulted in a Lyapunov function time derivative that was only negative semi-definite, but we eventually argued that the fixed points were asymptotically stable. That took a little extra work, involving an argument about the fact that the fixed points were the only place that the system could stay with $\dot{E}=0$; every other state with $\dot{E}=0$ was only transient. We can formalize this idea for the more general Lyapunov function statements--it is known as LaSalle's Theorem.

LaSalle's Theorem

Given a system $\dot{\bx} = f(\bx)$ with $f$ continuous. If we can produce a scalar function $V(\bx)$ with continuous derivatives for which we have $$V(\bx) \succ 0,\quad \dot{V}(\bx) \preceq 0,$$ and $V(\bx)\rightarrow \infty$ as $||\bx||\rightarrow \infty$, then $\bx$ will converge to the largest invariant set where $\dot{V}(\bx) = 0$.

To be clear, an invariant set, ${\cal G}$, of the dynamical system is a set for which $\bx(0)\in{\cal G} \Rightarrow \forall t>0, \bx(t) \in {\cal G}$. In other words, once you enter the set you never leave. The "largest invariant set" need not be connected; in fact for the pendulum example each fixed point is an invariant set, so the largest invariant set is the union of all the fixed points of the system. There are also variants of LaSalle's Theorem which work over a region.

Finding a Lyapunov function which $\dot{V} \prec 0$ is more difficult than finding one that has $\dot{V} \preceq 0$. LaSalle's theorem gives us the ability to make a statement about asymptotic stability even in this case. In the pendulum example, every state with $\dot\theta=0$ had $\dot{E}=0$, but only the fixed points are in the largest invariant set.

Swing-up for the Cart-Pole System

Recall the example of using partial-feedback linearization to generate a swing-up controller for the cart-pole system. We first examined the dynamics of the pole (pendulum) only, by writing it's energy: $$E(\bx) = \frac{1}{2}\dot\theta^2 - \cos\theta,$$ desired energy, $E^d = 1$, and the difference $\tilde{E}(\bx) = E(\bx) - E^d.$ We were able to show that our proposed controller produced $$\dot{\tilde{E}} = -k \dot\theta^2 \cos^2\theta \tilde{E},$$ where $k$ is a positive gain that we get to choose. And we said that was good!

Now we have the tools to understand that really we have a Lyapunov function $$V(\bx) = \frac{1}{2}\tilde{E}^2(\bx),$$, and what we have shown is that $\dot{V} \le 0$. By LaSalle, we can only argue that the closed-loop system will converge to the largest invariant set, which here is the entire homoclinic orbit: $\tilde{E}(\bx) = 0$. We have to switch to the LQR controller in order to stabilize the upright.

Lyapunov function: $V(\bx) = \frac{1}{2}\tilde{E}^2(\bx).$
Time-derivative of the Lyapunov function: $\dot{V}(\bx).$

As you can see from the plots, $\dot{V}(\bx)$ ends up being a quite non-trivial function! We'll develop the computational tools for verifying the Lyapunov/LaSalle conditions for systems of this complexity in the upcoming sections.

Relationship to the Hamilton-Jacobi-Bellman equations

At this point, you might be wondering if there is any relationship between Lyapunov functions and the cost-to-go functions that we discussed in the context of dynamic programming. After all, the cost-to-go functions also captured a great deal about the long-term dynamics of the system in a scalar function. We can see the connection if we re-examine the HJB equation \[ 0 = \min_\bu \left[ \ell(\bx,\bu) + \pd{J^*}{\bx}f(\bx,\bu). \right] \]Let's imagine that we can solve for the optimizing $\bu^*(\bx)$, then we are left with $ 0 = \ell(\bx,\bu^*) + \pd{J^*}{\bx}f(\bx,\bu^*) $ or simply \[ \dot{J}^*(\bx) = -\ell(\bx,\bu^*) \qquad \text{vs} \qquad \dot{V}(\bx) \preceq 0. \] In other words, in optimal control we must find a cost-to-go function which matches this gradient for every $\bx$; that's very difficult and involves solving a potentially high-dimensional partial differential equation. By contrast, Lyapunov analysis is asking for much less - any function which is going downhill (at any rate) for all states. This can be much easier, for theoretical work, but also for our numerical algorithms. Also note that if we do manage to find the optimal cost-to-go, $J^*(\bx)$, then it can also serve as a Lyapunov function so long as $\ell(\bx,\bu^*(\bx)) \succeq 0$.

Include instability results, as in Briat15 Theorem 2.2.5

Lyapunov analysis with convex optimization

One of the primary limitations in Lyapunov analysis as I have presented it so far is that it is potentially very difficult to come up with suitable Lyapunov function candidates for interesting systems, especially for underactuated systems. ("Underactuated" is almost synonymous with "interesting" in my vocabulary.) Even if somebody were to give me a Lyapunov candidate for a general nonlinear system, the Lyapunov conditions can be difficult to check -- for instance, how would I check that $\dot{V}$ is strictly negative for all $\bx$ except the origin if $\dot{V}$ is some arbitrarily complicated nonlinear function over a vector $\bx$?

In this section, we'll look at some computational approaches to verifying the Lyapunov conditions, and even to searching for (the coefficients of) the Lyapunov functions themselves.

If you're imagining numerical algorithms to check the Lyapunov conditions for complicated Lyapunov functions and complicated dynamics, the first thought is probably that we can evaluate $V$ and $\dot{V}$ at a large number of sample points and check whether $V$ is positive and $\dot{V}$ is negative. This does work, and could potentially be combined with some smoothness or regularity assumptions to generalize beyond the sample points. Add python bindings for the pendulum energy lp lyapunov example. But in many cases we will be able to do better -- providing optimization algorithms that will rigorously check these conditions for all $\bx$ without dense sampling; these will also give us additional leverage in formulating the search for Lyapunov functions.

Lyapunov analysis for linear systems

Let's take a moment to see how things play out for linear systems.

Lyapunov analysis for stable linear systems

Imagine you have a linear system, $\dot\bx = {\bf A}\bx$, and can find a Lyapunov function $$V(\bx) = \bx^T {\bf P} \bx, \quad {\bf P} = {\bf P^T} \succ 0,$$ which also satisfies $$\dot{V}(\bx) = \bx^T {\bf PA} \bx + \bx^T {\bf A}^T {\bf P}\bx \prec 0.$$ Then the origin is globally asymptotically stable.

Note that the radially-unbounded condition is satisfied by ${\bf P} \succ 0$, and that the derivative condition is equivalent to the matrix condition $${\bf PA} + {\bf A}^T {\bf P} \prec 0.$$

For stable linear systems the existence of a quadratic Lyapunov function is actually a necessary (as well as sufficient) condition. Furthermore, a Lyapunov function can always be found by finding the positive-definite solution to the matrix Lyapunov equation \begin{equation}{\bf PA} + {\bf A}^T{\bf P} = - {\bf Q},\label{eq:algebraic_lyapunov} \end{equation} for any ${\bf Q}={\bf Q}^T\succ 0$.

add an example here. double integrator? re-analyze the LQR output?

This is a very powerful result - for nonlinear systems it will be potentially difficult to find a Lyapunov function, but for linear systems it is straight-forward. In fact, this result is often used to propose candidates for non-linear systems, e.g., by linearizing the equations and solving a local linear Lyapunov function which should be valid in the vicinity of a fixed point.

Lyapunov analysis as a semi-definite program (SDP)

Lyapunov analysis for linear systems has an extremely important connection to convex optimization. In particular, we could have also formulated the Lyapunov conditions for linear systems above using semi-definite programming (SDP). Semidefinite programming is a subset of convex optimization -- an extremely important class of problems for which we can produce efficient algorithms that are guaranteed find the global optima solution (up to a numerical tolerance and barring any numerical difficulties).

If you don't know much about convex optimization or want a quick refresher, please take a few minutes to read the optimization preliminaries in the appendix. The main requirement for this section is to appreciate that it is possible to formulate efficient optimization problems where the constraints include specifying that one or more matrices are positive semi-definite (PSD). These matrices must be formed from a linear combination of the decision variables. For a trivial example, the optimization $$\min_a a,\quad \subjto \begin{bmatrix} a & 0 \\ 0 & 1 \end{bmatrix} \succeq 0,$$ returns $a = 0$ (up to numerical tolerances).

The value in this is immediate for linear systems. For example, we can formulate the search for a Lyapunov function for the linear system $\dot\bx = {\bf A} \bx$ by using the parameters ${\bf p}$ to populate a symmetric matrix ${\bf P}$ and then write the SDP: \begin{equation} \find_{\bf p} \quad \subjto \quad {\bf P} \succeq 0, \quad {\bf PA} + {\bf A}^T {\bf P} \preceq 0.\label{eq:lyap} \end{equation} Note that you would probably never use that particular formulation, since there specialized algorithms for solving the simple Lyapunov equation which are more efficient and more numerically stable. But the SDP formulation does provide something new -- we can now easily formulate the search for a "common Lyapunov function" for uncertain linear systems.

Common Lyapunov analysis for linear systems

Suppose you have a system governed by the equations $\dot\bx = {\bf A}\bx$, where the matrix ${\bf A}$ is unknown, but its uncertain elements can be bounded. There are a number of ways to write down this uncertainty set; let us choose to write this by describing ${\bf A}$ as the convex combination of a number of known matrices, $${\bf A} = \sum_{i} \beta_i {\bf A}_i, \quad \sum_i \beta_i = 1, \quad \forall i, \beta_i > 0.$$ This is just one way to specify the uncertainty; geometrically it is describing a polygon of uncertain parameters (in the space of elements of ${\bf A}$ with each ${\bf A}_i$ as one of the vertices in the polygon.

Now we can formulate the search for a common Lyapunov function using \[ \find_{\bf p} \quad \subjto \quad {\bf P} \succeq 0, \quad \forall_i, {\bf PA}_i + {\bf A}_i^T {\bf P} \preceq 0.\] The solver will then return a matrix ${\bf P}$ which satisfies all of the constraints, or return saying "problem is infeasible". It can easily be verified that if ${\bf P}$ satisfies the Lyapunov condition at all of the vertices, then it satisfies the condition for every ${\bf A}$ in the set: \[ {\bf P}(\sum_i \beta_i {\bf A}_i) + (\sum_i \beta_i {\bf A}_i)^T {\bf P} = \sum_i \beta_i ({\bf P A}_i + {\bf A}_i^T {\bf P}) \preceq 0, \] since $\forall i$, $\beta_i > 0$. Note that, unlike the simple Lyapunov equation for a known linear system, this condition being satisfied is a sufficient but not a necessary condition -- it is possible that the set of uncertain matrices ${\bf A}$ is robustly stable, but that this stability cannot be demonstrated with a common quadratic Lyapunov function.

You can try this example for yourself in .

As always, make sure that you open up the code and take a look.

There are many small variants of this result that are potentially of interest. For instance, a very similar set of conditions can certify "mean-square stability" for linear systems with multiplicative noise (see e.g. Boyd94, § 9.1.1).

Add example or exercise based on e.g. Briat15 (LPV) example 1.3.1, showing off how powerful this can be. Also perhaps the parameter-dependent robust stability of his Def 2.3.5.

This example is very important because it establishes a connection between Lyapunov functions and (convex) optimization. But so far we've only demonstrated this connection for linear systems where the PSD matrices provide a magical recipe for establishing the positivity of the (quadratic) functions for all $\bx$. Is there any hope of extending this type of analysis to more general nonlinear systems? Surprisingly, it turns out that there is.

Lyapunov analysis for polynomial systems

Sums of squares optimization provides a natural generalization of SDP to optimizing over positive polynomials (if you are not familiar, take a moment to read the appendix). This suggests that it may be possible to generalize the optimization approach using SDP to search for Lyapunov functions for linear systems to searching for Lyapunov functions for at least the polynomial systems: $\dot\bx = f(\bx),$ where $f$ is a vector-valued polynomial function. If we parameterize a fixed-degree Lyapunov candidate as a polynomial with unknown coefficients, e.g., \[ V_\alpha(\bx) = \alpha_0 + \alpha_1 x_1 + \alpha_2 x_2 + \alpha_3 x_1x_2 + \alpha_4 x_1^2 + ..., \] then the time-derivative of $V$ is also a polynomial, and I can formulate the optimization: \begin{align*} \find_\alpha, \quad \subjto \quad& V_\alpha(\bx) \sos \\ & -\dot{V}_\alpha(\bx) = -\pd{V_\alpha}{\bx} f(\bx) \sos. \end{align*} Because this is a convex optimization, the solver will return a solution if one exists.

Verifying a Lyapunov candidate via SOS

This example is example 7.2 from Parrilo00. Consider the nonlinear system: \begin{align*} \dot{x}_0 =& -x_0 - 2x_1^2 \\ \dot{x}_1 =& -x_1 - x_0 x_1 - 2x_1^3,\end{align*} and the fixed Lyapunov function $V(x) = x_0^2 + 2x_1^2$, test if $\dot{V}(x)$ is negative definite.

The numerical solution can be written in a few lines of code, and is a convex optimization.

Cart-pole swingup (again)

In the cart-pole swingup example, we took $$V(\bx) = \frac{1}{2} \tilde{E}^2(\bx) = \frac{1}{2} (\frac{1}{2}\dot\theta^2 - \cos\theta - 1)^2.$$ This is clearly a sum of squares. Furthermore, we showed that $$\dot{V}(\bx) = - \dot\theta^2 \cos^2\theta \tilde{E}^2(\bx),$$ which is also a sum of squares. So the proposal of using sums-of-squares optimization is not so different, actually, than the recipes that nonlinear control theorists have been using (on pen and paper) for years. In this case, I would need a basis vector that includes many more monomials: $[1, \dot\theta, \cos\theta, ..., \dot\theta^3, ... ]^T,$ and setting up the equality constraints in an optimization requires more complicated term matching when trigonometric functions are involved, but the recipe still works.

Searching for a Lyapunov function via SOS

Verifying a candidate Lyapunov function is all well and good, but the real excitement starts when we use optimization to find the Lyapunov function. In the following code, we parameterize $V(x)$ as the polynomial containing all monomials up to degree 2, with the coefficients as decision variables: $$V(x) = c_0 + c_1x_0 + c_2x_1 + c_3x_0^2 + c_4 x_0x_1 + c_5 x_1^2.$$ We will set the scaling (arbitrarily) to avoid numerical issues by setting $V(0)=0$, $V([1,0]) = 1$. Then we write: \begin{align*} \find_{\bc} \ \ \subjto \ \ & V\text{ is sos, } \\ & -\dot{V} \text{ is sos.}\end{align*}

Up to numerical convergence tolerance, it discovers the same coefficients that we chose above (zeroing out the unnecessary terms).

It is important to remember that there are a handful of gaps which make the existence of this solution a sufficient condition (for proving that every sublevel set of $V$ is an invariant set of $f$) instead of a necessary one. First, there is no guarantee that a stable polynomial system can be verified using a polynomial Lyapunov function (of any degree, and in fact there are known counter-examples Ahmadi11a) and here we are only searching over a fixed-degree polynomial. Second, even if a polynomial Lyapunov function does exist, there is a gap between the SOS polynomials and the positive polynomials.

Despite these caveats, I have found this formulation to be surprisingly effective in practice. Intuitively, I think that this is because there is relatively a lot of flexibility in the Lyapunov conditions -- if you can find one function which is a Lyapunov function for the system, then there are also many "nearby" functions which will satisfy the same constraints.

Lyapunov functions for estimating regions of attraction

There is another very important connection between Lyapunov functions and the concept of an invariant set: any sublevel set of a Lyapunov function is also an invariant set. This gives us the ability to use sublevel sets of a Lyapunov function as approximations of the region of attraction for nonlinear systems.

Lyapunov invariant set and region of attraction theorem

Given a system $\dot{\bx} = f(\bx)$ with $f$ continuous, if we can find a scalar function $V(\bx) \succ 0$ and a sublevel set $${\cal G}: \{ \bx | V(\bx) \le \rho \}$$ on which $$\forall \bx \in {\cal G}, \dot{V}(\bx) \preceq 0,$$ then ${\cal G}$ is an invariant set. By LaSalle, $\bx$ will converge to the largest invariant subset of ${\cal G}$ on which $\dot{V}=0$.

Furthermore, if $\dot{V}(\bx) \prec 0$ in ${\cal G}$, then the origin is locally asymptotically stable and the set ${\cal G}$ is inside the region of attraction of this fixed point. Alternatively, if $\dot{V}(\bx) \preceq 0$ in ${\cal G}$ and $\bx = 0$ is the only invariant subset of ${\cal G}$ where $\dot{V}=0$, then the origin is asymptotically stable and the set ${\cal G}$ is inside the region of attraction of this fixed point.

Region of attraction for a one-dimensional system

Consider the first-order, one-dimensional system $\dot{x} = -x + x^3.$ We can quickly understand this system using our tools for graphical analysis.

In the vicinity of the origin, $\dot{x}$ looks like $-x$, and as we move away it looks increasingly like $x^3$. There is a stable fixed point at the origin and unstable fixed points at $\pm 1$. In fact, we can deduce visually that the region of attraction to the stable fixed point at the origin is $\bx \in (-1,1)$. Let's see if we can demonstrate this with a Lyapunov argument.

First, let us linearize the dynamics about the origin and use the Lyapunov equation for linear systems to find a candidate $V(\bx)$. Since the linearization is $\dot{x} = -x$, if we take ${\bf Q}=1$, then we find ${\bf P}=\frac{1}{2}$ is the positive definite solution to the algebraic Lyapunov equation (\ref{eq:algebraic_lyapunov}). Proceeding with $$V(\bx) = \frac{1}{2} x^2,$$ we have $$\dot{V} = x (-x + x^3) = -x^2 + x^4.$$ This function is zero at the origin, negative for $|x| < 1$, and positive for $|x| > 1$. Therefore we can conclude that the sublevel set $V < \frac{1}{2}$ is invariant and the set $x \in (-1,1)$ is inside the region of attraction of the nonlinear system. In fact, this estimate is tight.

Robustness analysis using "common Lyapunov functions"

While we will defer most discussions on robustness analysis until later in the notes, the idea of a common Lyapunov function, which we introduced briefly for linear systems in the example above, can be readily extended to nonlinear systems and region of attraction analysis. Imagine that you have a model of a dynamical system but that you are uncertain about some of the parameters. For example, you have a model of a quadrotor, and are fairly confident about the mass and lengths (both of which are easy to measure), but are not confident about the moment of inertia. One approach to robustness analysis is to define a bounded uncertainty, which could take the form of $$\dot{\bx} = f_\alpha(\bx), \quad \alpha_{min} \le \alpha \le \alpha_{max},$$ with $\alpha$ a vector of uncertain parameters in your model. Richer specifications of the uncertainty bounds are also possible, but this will serve for our examples.

In standard Lyapunov analysis, we are searching for a function that goes downhill for all $\bx$ to make statements about the long-term dynamics of the system. In common Lyapunov analysis, we can make many similar statements about the long-term dynamics of an uncertain system if we can find a single Lyapunov function that goes downhill for all possible values of $\alpha$. If we can find such a function, then we can use it to make statements with all of the variations we've discussed (local, regional, or global; in the sense of Lyapunov, asymptotic, or exponential).

A one-dimensional system with gain uncertainty

Let's consider the same one-dimensional example used above, but add an uncertain parameter into the dynamics. In particular, consider the system: $$\dot{x} = -x + \alpha x^3, \quad \frac{3}{4} < \alpha < \frac{3}{2}.$$ Plotting the graph of the one-dimensional dynamics for a few values of $\alpha$, we can see that the fixed point at the origin is still stable, but the robust region of attraction to this fixed point (shaded in blue below) is smaller than the region of attraction for the system with $\alpha=1$.

Taking the same Lyapunov candidate as above, $V = \frac{1}{2} x^2$, we have $$\dot{V} = -x^2 + \alpha x^4.$$ This function is zero at the origin, and negative for all $\alpha$ whenever $x^2 > \alpha x^4$, or $$|x| < \frac{1}{\sqrt{\alpha_{max}}} = \sqrt{\frac{2}{3}}.$$ Therefore, we can conclude that $|x| < \sqrt{\frac{2}{3}}$ is inside the robust region of attraction of the uncertain system.

Not all forms of uncertainty are as simple to deal with as the gain uncertainty in that example. For many forms of uncertainty, we might not even know the location of the fixed points of the uncertain system. In this case, we can often still use common Lyapunov functions to give some guarantees about the system, such as guarantees of robust set invariance. For instance, if you have uncertain parameters on a quadrotor model, you might be ok with the quadrotor stabilizing to a pitch of $0.01$ radians, but you would like to guarantee that it definitely does not flip over and crash into the ground.

A one-dimensional system with additive uncertainty

Now consider the system: $$\dot{x} = -x + x^3 + \alpha, \quad -\frac{1}{4} < \alpha < \frac{1}{4}.$$ Plotting the graph of the one-dimensional dynamics for a few values of $\alpha$, this time we can see that the fixed point is not necessarily at the origin; the location of the fixed point moves depending on the value of $\alpha$. But we should be able to guarantee that the uncertain system will stay near the origin if it starts near the origin, using an invariant set argument.

Taking the same Lyapunov candidate as above, $V = \frac{1}{2} x^2$, we have $$\dot{V} = -x^2 + x^4 + \alpha x.$$ This Lyapunov function allows us to easily verify, for instance, that $V \le \frac{1}{3}$ is a robust invariant set, because whenever $V = \frac{1}{3}$, we have $$\forall \alpha \in [\alpha_{min},\alpha_{max}],\quad \dot{V}(x,\alpha) < 0.$$ Therefore $V$ can never start at less than one-third and cross over to greater than one-third. To see this, see that $$ V=\frac{1}{3} \quad \Rightarrow \quad x = \pm \sqrt{\frac{2}{3}} \quad \Rightarrow \quad \dot{V} = -\frac{2}{9} \pm \alpha \sqrt{\frac{2}{3}} < 0, \forall \alpha \in \left[-\frac{1}{4},\frac{1}{4} \right]. $$ Note that not all sublevel sets of this invariant set are invariant. For instance $V < \frac{1}{32}$ does not satisfy this condition, and by visual inspection we can see that it is in fact not robustly invariant.

robust quadrotor example

Region of attraction estimation for polynomial systems

Now we have arrived at the tool that I believe can be a work-horse for many serious robotics applications. Most of our robots are not actually globally stable (that's not because they are robots -- if you push me hard enough, I will fall down, too), which means that understanding the regions where a particular controller can be guaranteed to work can be of critical importance.

Sums-of-squares optimization effectively gives us an oracle which we can ask: is this polynomial positive for all $\bx$? To use this for regional analysis, we have to figure out how to modify our questions to the oracle so that the oracle will say "yes" or "no" when we ask if a function is positive over a certain region which is a subset of $\Re^n$. That trick is called the S-procedure. It is closely related to the Lagrange multipliers from constrained optimization, and has deep connections to "Positivstellensatz" from algebraic geometry.

The S-procedure

Consider a scalar polynomial, $p(\bx)$, and a semi-algebraic set $g(\bx) \preceq 0$, where $g$ is a vector of polynomials. If I can find a polynomial "multiplier", $\lambda(\bx)$, such that \[ p(\bx) + \lambda^T(\bx) g(\bx) \sos, \quad \text{and} \quad \lambda(\bx) \sos, \] then this is sufficient to demonstrate that $$p(\bx)\ge 0 \quad \forall \bx \in \{ \bx | g(\bx) \le 0 \}.$$ To convince yourself, observe that when $g(\bx) \le 0$, it is only harder to be positive, but when $g(\bx) > 0$, it is possible for the combined function to be SOS even if $p(\bx)$ is negative. We will sometimes find it convenient to use the short-hand: $$g(\bx) \le 0 \Rightarrow p(\bx) \ge 0$$ to denote the implication certified by the S-procedure (e.g. "whenever $g(\bx) \le 0$, we have $p(\bx) \ge 0$").

We can also handle equality constraints with only a minor modification -- we no longer require the multiplier to be positive. If I can find a polynomial "multiplier", $\lambda(\bx)$, such that \[p(\bx) + \lambda^T(\bx) g(\bx) \sos \] then this is sufficient to demonstrate that $$p(\bx)\ge 0 \quad \forall \bx \in \{ \bx | g(\bx) = 0 \}.$$ Here the intuition is that $\lambda(x)$ can add arbitrary positive terms to help me be SOS, but those terms contribute nothing precisely when $g(x)=0$.

Basic region of attraction formulation

The S-procedure gives us the tool we need to evaluate positivity only over a region of state space, which is precisely what we need to certify the Lyapunov conditions for a region-of-attraction analysis. Let us start with a positive-definite polynomial Lyapunov candidate, $V(\bx) \succ 0$, then we can write the Lyapunov conditions: $$\dot{V}(\bx) \prec 0 \quad \forall \bx \in \{ \bx | V(\bx) \le \rho \},$$ using sums-of-squares and the S-procedure: $$-\dot{V}(\bx) + \lambda(\bx)(V(\bx) - \rho) \text{ is SOS,} \quad \text{and} \quad \lambda(\bx) \text{ is SOS},$$ where $\lambda(\bx)$ is a multiplier polynomial with free coefficients that are to be solved for in the optimization.

I think it's easiest to see the details in an example.

Region of attraction for the one-dimensional cubic system

Let's return to our example from above: \[ \dot{x} = -x + x^3 \] and try to use SOS optimization to demonstrate that the region of attraction of the fixed point at the origin is $x \in (-1,1)$, using the Lyapunov candidate $V = x^2.$

First, define the multiplier polynomial, \[ \lambda(x) = c_0 + c_1 x + c_2 x^2. \] Then define the optimization \begin{align*} \find_{\bf c} \quad & \\ \subjto \quad& - \dot{V}(x) + \lambda(x) (V(x)-1) \sos \\ & \lambda(x) \sos \end{align*}

You can try this example for yourself in .

In this example, we only verified that the one-sublevel set of the pre-specified Lyapunov candidate is negative (certifying the ROA that we already understood). Even more useful is if you are able to search for the largest $\rho$ that can satisfy these conditions. Unfortunately, in this first formulation, optimizing $\rho$ directly would make the optimization problem non-convex because we would have terms like $\rho c_0$, $\rho c_1 x$, ... which are bilinear in the decision variables; we need the sums-of-squares constraints to be only linear in the decision variables.

Fortunately, because the problem is convex with $\rho$ fixed (and therefore can be solved reliably), and $\rho$ is just a scalar, we can perform a simple line search on $\rho$ to find the largest value for which the convex optimization returns a feasible solution. This will be our estimate for the region of attraction.

There are a number of variations to this basic formulation; I will describe a few of them below. There are also important ideas like rescaling and degree-matching that can have a dramatic effect on the numerics of the problem, and potentially make them much better for the solvers. But you do not need to master them all in order to use the tools effectively.

Region of Attraction codes in Drake

In , we have packaged most of the work in setting up and solving the sums-of-squares optimization for regions of attraction into a single method RegionOfAttraction(system, context, options). This makes it as simple as, for instance:

x = Variable("x")
sys = SymbolicVectorSystem(state=[x], dynamics=[-x+x**3])
context = sys.CreateDefaultContext()
V = RegionOfAttraction(sys, context)

Remember that although we have tried to make it convenient to call these functions, they are not a black box. I highly recommend opening up the RegionOfAttraction method and understanding how it works. There are lots of different options / formulations, and numerous numerical recipes to improve the numerics of the optimization problem.

The equality-constrained formulation

Here is one important variation for finding the level set of a candidate region of attraction, which turns an inequality in the S-procedure into an equality. This formulation is jointly convex in $\lambda(\bx)$ and $\rho$, so one can optimize them in a single convex optimization. It appeared informally in an example in Parrilo00, and was discussed a bit more in Shen20.

Under the assumption that the Hessian of $\dot{V}(\bx)$ is negative-definite at the origin (which is easily checked), we can write \begin{align*} \max_{\rho, \lambda(\bx)} & \quad \rho \\ \subjto & \quad (\bx^T\bx)^d (V(\bx) - \rho) + \lambda(\bx) \dot{V}(\bx) \text{ is SOS},\end{align*} with $d$ a fixed positive integer. As you can see, $\rho$ no longer multiplies the coefficient of $\lambda(\bx)$. But why does this certify a region of attraction?

You can read the sums-of-squares constraint as certifying the implication that whenever $\dot{V}(x) = 0$, we have that either $V(\bx) \ge \rho$ OR $\bx = 0$. Multiplying by some multiple of $\bx^T\bx$ is just a clever way to handle the "or $\bx=0$" case, which is necessary since we expect $\dot{V}(0) = 0$. This implication is sufficient to prove that $\dot{V}(\bx) \le 0$ whenever $V(\bx) \le \rho$, since $V$ and $\dot{V}$ are smooth polynomials; we examine this claim in full detail in one of the exercises.

Using the S-procedure with equalities instead of inequalities also has the potential advantage of removing the SOS constraint on $\lambda(\bx).$ But perhaps the biggest advantage of this formulation is the possibility of dramatically simplifying the problem using the quotient ring of this algebraic variety, and in particular some recent results for exact certification using sampling varietiesShen20.

Searching for $V(\bx)$

The machinery so far has used optimization to find the largest region of attraction that can be certified given a candidate Lyapunov function. This is not necessarily a bad assumption. For most stable fixed points, we can certify the local stability with a linear analysis, and this linear analysis gives us a candidate quadratic Lyapunov function that can be used for nonlinear analysis over a region. Practically speaking, when we solve an LQR problem, the cost-to-go from LQR is a good candidate Lyapunov function for the nonlinear system. If we are simply analyzing an existing system, then we can obtain this candidate by solving a Lyapunov equation (Eq \ref{eq:algebraic_lyapunov}).

But what if we believe the system is regionally stable, despite having an indefinite linearization? Or perhaps we can certify a larger volume of state space by using a Lyapunov candidate with degree greater than two. Can we use sums-of-squares optimization to find that in the region of attraction case, too?

To accomplish this, we will now make $V(x)$ a polynomial (of some fixed degree) with the coefficients as decision variables. First we will need to add constraints to ensure that $$V(0) = 0 \quad {and} \quad V(\bx) - \epsilon \bx^T\bx \text{ is SOS},$$ where $\epsilon$ is some small positive constant. This $\epsilon$ term simply ensures that $V$ is strictly positive definite. Now let's consider our basic formulation: $$-\dot{V}(\bx) + \lambda(\bx)(V(\bx) - 1) \text{ is SOS,} \quad \text{and} \quad \lambda(\bx) \text{ is SOS}.$$ Notice that I've replaced $\rho=1$; now that we are searching for $V$ the scale of the level-set can be set aribtrarily to 1. In fact, it's better to do so -- if we did not set $V(0)=0$ and $V(\bx) \le 1$ as the sublevel set, then the optimization problem would be underconstrained, and might cause problems for the solver.

Unfortunately, the derivative constraint is now nonconvex (bilinear) in the decision variables, since we are searching for both the coefficients of $\lambda$ and $V$, and they are multiplied together. Our equality-constrained formulation doesn't get us around this one, either (since the coefficients of $V$ also appear in $\dot{V}$). Here we have to resort to some weaker form of optimization. In practice, we have had good practical success using bilinear alternations: start with an initial $V$ (e.g. from LQR), and search for $\lambda$; then fix $\lambda$ and search for $V$ and repeat until convergence (see, for instance Tedrake10+Majumdar16a). A typical choice for the objective is to maximize some convex surrogate for the volume of the certified region of attraction in state space; typically we use the determinant of the quadratic form describing a contained ellipse.

Barring numerical issues, this algorithm is guaranteed to have recursive feasibility and tends to converge in just a few alternations, but there is no guarantee that it finds the optimal solution.

Searching for Lyapunov functions

(Details coming soon...)

Convex outer approximations of the ROA

All of our region of attraction approximations so far have been "inner approximations" -- the certified region is guaranteed to be contained within the true region of attraction for the system. This is what we want if we are to give some guarantee of stability. As we increase the degree of our polynomial and multipliers, we expect that the estimated regions of attraction will converge on the true regions.

It turns out that if we instead consider outer approximations, which converge on the true ROA from the other side, we can write formulations that enable searching for the Lyapunov function directly as a convex optimization. As we will see below, this approach also allows one to write convex formulations for controller design. These methods have been explored in a very nice line of work by Henrion and Korda (e.g. Henrion14), using the method of moments from Lasserre10 also called "occupation measures". Their treatment emphasizes infinite-dimensional linear programs and heirarchies of LMI relaxations; these are just dual formulations to the SOS optimizations that we've been writing here. I find the SOS form more clear, so will stick to it here. Some of the occupation measure papers include the dual formulation of the infinite-dimension linear program which will look quite similar; my coauthors and I typically call out the sums-of-squares version in our papers on the topic (e.g. Posa17).

To find an outer approximation, instead of solving for a Lyapunov function that certifies convergence to the origin, we use a very related set of conditions to search for a Lyapunov-like "barrier certificate", $\mathcal{B}(\bx).$ Like a Lyapunov function, we'd like $\dot{\mathcal{B}}(\bx) \leq 0$; this time we'll ask for this to be true everywhere (or at least in some set that is sufficiently larger than the ROA of interest). Then we will set $\mathcal{B}(0) > 0$. If we can find such a function, then certainly any state for which $\mathcal{B}(\bx) < 0$ is outside of the region of attraction of the fixed point -- since $\mathcal{B}$ cannot increase it can never reach the origin which has $\mathcal{B} > 0$. The superlevel set $\{ \bx | \mathcal{B}(\bx) \ge 0\}$ is an outer approximation of the true region of attraction: $$-\dot{\mathcal{B}}(\bx) \text{ is SOS} \quad \text{and} \quad \mathcal{B}(0) > 0.$$

In order to find the smallest such outer approximation (given a fixed degree polynomial), we choose an objective that tries to "push-down" on the value of $\mathcal{B}(\bx)$. We typically accomplish this by introducing another polynomial function $W(\bx)$ with the requirements that $$W(\bx) \ge 0 \quad \text{and} \quad W(\bx) \ge \mathcal{B}(\bx) +1,$$ implemented as SOS constraints. Then we minimize the integral $\int_\bx W(\bx)d\bx$, which can be readily computed over a compact set like a ball in $\Re^n$ Folland01 and is linear in the coefficients. More sophisticated alternatives also existHenrion09.

Outer approximation for the cubic polynomial

Let's revisit my favorite simple dynamical system, $\dot{x} = -x + x^3,$ which has a region of attraction for the fixed point at zero of $\{x \| |x|< 1 \}.$ This time, we'll estimate the ROA using the outer approximation. We can accomplish this with the following program: \begin{align*} \min_{\mathcal{B}(x), W(x)} \quad & \int_{-2}^2 W(x)dx, \\ \subjto \quad & -\dot{\mathcal{B}}(x) & \text{is SOS}, \\ & W(x) &\text{ is SOS}, \\ & W(x) - \mathcal{B}(x) - 1.0 & \text{ is SOS}, \\ & \mathcal{B}(0) \ge 0.\end{align*}

To make the problem a little numerically better, you'll see in the code that I've asked for $\dot{\mathcal{B}}(x)$ to be strictly negative definite, for $\mathcal{B}(0) \ge 0.1$, and I've chosen to only include even-degree monomials in $\mathcal{B}(x)$ and $W(x)$. Plotting the solution reveals:

As you can see, the superlevel set, $\mathcal{B}(x) \ge 0$ is a tight outer-approximation of the true region of attraction. In the limit of increasing the degrees of the polynomials to infinity, we would expect that $W(\bx)$ would converge to the indicator function that is one inside the region of attraction, and zero outside (we are quite far from that here, but nevertheless have a tight approximation from $\mathcal{B}(\bx) \ge 0$).

Finite-time Reachability

So far we have used Lyapunov analysis to analyze stability, which is fundamentally a description of the system's behavior as time goes to infinity. But Lyapunov analysis can also be used to analyze the finite-time behavior of nonlinear systems. We will see a number of applications of this finite-time analysis over the next few chapters. It can be applied even to unstable systems, or systems that are stable to a limit cycle instead of a fixed-point. It can also be applied to systems that are only defined over a finite interval of time, as might be the case if we are executing a planned trajectory with a bounded duration.

Time-varying dynamics and Lyapunov functions

Before focusing on the finite time, let us first realize that the basic (infinite-time) Lyapunov analysis can also be applied to time-varying systems: $\dot{\bx} = f(t, \bx).$ We can analyze the stability (local, regional, or global) of this system with very little change. If we find \begin{gather*}V(\bx) \succ 0, \\ \forall t,\forall \bx \ne 0, \quad \dot{V}(t, \bx) = \pd{V}{\bx}f(t,\bx) < 0, \quad \dot{V}(t, 0) = 0, \end{gather*} then all of our previous statement still hold. In our SOS formulations, $t$ is simply one more indeterminate.

Similarly, even for a time-invariant system, it is also possible to define a time-varying Lyapunov function, $V(t, \bx)$ and establish local, regional, or global stability using the almost the same conditions: \begin{gather*}\forall t, \forall \bx \ne 0, \quad V(t, \bx) > 0, \quad V(t, 0) = 0, \\ \forall t, \forall \bx \ne 0, \quad \dot{V}(t, \bx) = \pd{V}{\bx}f(\bx) + \pd{V}{t} < 0, \quad \dot{V}(t, 0) = 0. \end{gather*}

These two ideas each stand on their own, but they very often go together, as time-varying dynamics are perhaps the best motivator for studying a time-varying Lyapunov function. Afterall, we do know that a stable time-invariant system must have a time-invariant Lyapunov function that demonstrates this stability (from the "converse Lyapunov function" theorems). But we do not know apriori how to represent this function; as an example, remember we know that there are stable polynomial systems that cannot be verified with a polynomial Lyapunov function. For global stability analysis, the time-varying Lyapunov analysis does not add modeling power: since the conditions must be satisfied for all $t$, we could have just set $t$ to a constant and used the time-invariant function. But there may be cases where a time-varying analysis could afford a different analysis for the regional or local stability cases. We will see a good example of this when we study limit cycles.

Finite-time reachability

Finite-time reachability analysis is an important concept for control design and verification, where we seek to understand the behavior of a system over only a finite time interval, $[t_1, t_2]$. It is almost always a region-based analysis, and attempts to make a statement of the form: $$\bx(t_1) \in \mathcal{X}_1 \Rightarrow \bx(t_2) \in \mathcal{X}_2,$$ where $\mathcal{X}_1, \mathcal{X}_2 \subset \Re^n$ are regions of state space. More generally, we might like to understand the time-varying reachable set $\forall t\in [t_1, t_2], \mathcal{X}(t).$

Reachability analysis can be done forward in time: we choose $\mathcal{X}_1$ and try to find the smallest region $\mathcal{X}_2$ for which we can make the statement hold. $\mathcal{X}(t)$ would be called the forward-reachable set (FRS), and can be very useful for certifying e.g. a motion plan. For instance, you might like to prove that your UAV does not crash into a tree in the next 5 seconds. In this case $\mathcal{X}_1$ might be take to be a point representing the current state of the vehicle, or a region representing an outer-approximation of the current state if the state in uncertain. In this context, we would call $\mathcal{X}_2$ a forward-reachable set. In this use case, we would typically choose any approximations in our Lyapunov analysis to certify that an estimate of the reachable region is also an outer-approximation: $\mathcal{X}_2 \subseteq \hat{\mathcal{X}}_2.$

Reachability analysis can also be done backward in time: we choose $\mathcal{X}_2$ and try to maximize the region $\mathcal{X}_1$ for which the statement can be shown to hold. Now $\mathcal{X}(t)$ is called the backward-reachable set (BRS), and for robustness we typically try to certify that our estimates are an inner-approximation, $\hat{\mathcal{X}}_1 \subseteq \mathcal{X}_1.$ The region-of-attraction analysis we studied above can be viewed as a special case of this, with $\mathcal{X}_2$ taken to be the fixed-point, $t_2 = 0$ and $t_1 = -\infty$. But finite-time BRS also have an important role to play, for instance when we are composing multiple controllers in order to achieve a more complicated task, which we will study soon.

Reachability via Lyapunov functions

Lyapunov functions can be used to certify finite-time reachability, even for continuous-time systems. The basic recipe is to certify the Lyapunov conditions over a (potentially time-varying) invariant set. Once again, we typically represent this as a (time-varying) level set of the Lyapunov function containing the origin, $V(\bx) \le \rho(t),$ where $\rho(t)$ is now a positive scalar function of time. Since we already have time as a decision variable, we can easily accommodate time-varying dynamics and Lyapunov functions, as well, so: \begin{align*} \forall t\in[t_1, t_2], \quad \forall \bx, \qquad& V(t,\bx) > 0, \quad& V(t, 0) = 0 \\ \forall t \in [t_1, t_2], \quad \forall \bx \in \{\bx | V(t,\bx) = \rho(t)\}, \qquad& \dot{V}(t, x) < \dot\rho(t).\end{align*} Note that finite-time reachability is about proving invariance of the set, not stability, and the $\dot{V}$ condition need only be certified at the boundary of the level set. If $V$ is decreasing at the boundary, then trajectories can never leave. One can certainly ask for more -- we may want to show that the system is converging towards $V(t,\bx) = 0$, perhaps even at some rate -- but only invariance is required to certify reachability.

Again, for polynomial systems and dynamics, sums-of-squares optimization is a powerful tool for certifying these properties numerically, and optimizing the volume of the estimated regions. Having taken $t$ as an indeterminate, we can use the S-procedure again to certify the conditions $\forall t \in [t_1, t_2].$

Like in the case for region of attraction, we have many formulations. We can certify an existing Lyapunov candidate, $V(t,\bx)$, and just try to maximize/minimize $\rho(t)$. Or we can search for the parameters of $V(t,\bx)$, too. Again, we can initialize that search using the time-varying version of the Lyapunov equation, or the solutions to a time-varying LQR Riccati equation.

In practice, we often certify the Lyapunov conditions over $\bx$ at only a finite set of samples $t_i \in [t_1, t_2]$. I don't actually have anything against sampling in one dimension; there are no issues with scaling to higher dimensions, and one can make practical rigorous statement about bounding the sampling error. And in these systems, adding $t$ into all of the equation otherwise can dramatically increase the degree of the polynomials required for the SOS certificates. All of this was written up nicely in Tobenkin10b, and robust variants of it were developed in Majumdar13f+Majumdar16a.

Rigid-body dynamics are (rational) polynomial

We've been talking a lot in this chapter about numerical methods for polynomial systems. But even our simple pendulum has a $\sin\theta$ in the dynamics. Have I been wasting your time? Must we just resort to polynomial approximations of the non-polynomial equations? It turns out that our polynomial tools can perform exact analysis of the manipulation equation for almost all††the most notable exception to this is if your robot has helical/screw joints (see Wampler11).of our robots. We just have to do a little more work to reveal that structure.

Let us first observe that rigid-body kinematics are polynomial (except the helical joint). This is fundamental -- the very nature of a "rigid body" assumption is that Euclidean distance is preserved between points on the body; if $\bp_1$ and $\bp_2$ are two points on a body, then the kinematics enforce that $|\bp_1 - \bp_2|_2^2$ is constant -- these are polynomial constraints. Of course, we commonly write the kinematics in a minimal coordinates using $\sin\theta$ and $\cos\theta$. But because of rigid body assumption, these terms only appear in the simplest forms, and we can simply make new variables $s_i = \sin\theta_i, c_i = \cos\theta_i$, and add the constraint that $s_i^2 + c_i^2 = 1.$ For a more thorough discussion see, for instance, Wampler11 and Sommese05. Since the potential energy of a multi-body system is simply an accumulation of weight times the vertical position for all of the points on the body, the potential energy is polynomial.

If configurations (positions) of our robots can be described by polynomials, then velocities can as well: forward kinematics $\bp_i = f(\bq)$ implies that $\dot\bp_i = \frac{\partial f}{\partial \bq}\dot\bq,$ which is polynomial in $s, c, \dot\theta$. Since the kinetic energy of our robot is given by the accumulation of the kinetic energy of all the mass, $T = \sum_i \frac{1}{2} m_i v_i^Tv_i,$ the kinetic energy is polynomial, too (even when we write it with inertial matrices and angular velocities).

Finally, the equations of motion can be obtained by taking derivatives of the Lagrangian (kinetic minus potential). These derivatives are still polynomial!

Global stability of the simple pendulum via SOS

We opened this chapter using our intuition about energy to discuss stability on the simple pendulum. Now we'll replace that intuition with convex optimization (because it will also work for more difficult systems where our intuition fails).

Let's change coordinates from $[\theta,\dot\theta]^T$ to $\bx = [s,c,\dot\theta]^t$, where $s \equiv \sin\theta$ and $c \equiv \cos\theta$. Then we can write the pendulum dynamics as $$\dot\bx = \begin{bmatrix} c \dot\theta \\ -s \dot\theta \\ -\frac{1}{m l^2} \left( b \dot\theta + mgls \right) \end{bmatrix}.$$

Now let's parameterize a Lyapunov candidate $V(s,c,\dot\theta)$ as the polynomial with unknown coefficients which contains all monomials up to degree 2: $$V = \alpha_0 + \alpha_1 s + \alpha_2 c + ... \alpha_{9} s^2 + \alpha_{10} sc + \alpha_{11} s\dot\theta.$$ Now we'll formulate the feasibility problem: \[ \find_{\bf \alpha} \quad \subjto \quad V \sos, \quad -\dot{V} \sos.\] In fact, this is asking too much -- really $\dot{V}$ only needs to be negative when $s^2+c^2=1$. We can accomplish this with the S-procedure, and instead write \[ \find_{{\bf \alpha},\lambda} \quad \subjto \quad V \sos, \quad -\dot{V} -\lambda(\bx)(s^2+c^2-1) \sos.\] (Recall that $\lambda(\bx)$ is another polynomial with free coefficients which the optimization can use to make terms arbitrarily more positive when $s^2+c^2 \neq 1$.) Finally, for style points, in the code example in we ask for exponential stability:

As always, make sure that you open up the code and take a look. The result is a Lyapunov function that looks familiar (visualized as a contour plot here):

Aha! Not only does the optimization finds us coefficients for the Lyapunov function which satisfy our Lyapunov conditions, but the result looks a lot like mechanical energy. In fact, the result is a little better than energy... there are some small extra terms added which prove exponential stability without having to invoke LaSalle's Theorem.

The one-degree-of-freedom pendulum did allow us to gloss over one important detail: while the manipulator equations $\bM(\bq) \ddot{\bq} + \bC(\bq, \dot\bq)\dot{\bq} = ...$ are polynomial, in order to solve for $\ddot{\bq}$ we actually have to multiply both sides by $\bM^{-1}$. This, unfortunately, is not a polynomial operation, so in fact the final dynamics of the multibody systems are rational polynomial. Not only that, but evaluating $\bM^{-1}$ symbolically is not advised -- the equations get very complicated very fast. But we can actually write the Lyapunov conditions using the dynamics in implicit form, e.g. by writing $V(\bq,\dot\bq,\ddot\bq)$ and asking it to satisfy the Lyapunov conditions everywhere that $\bM(\bq)\ddot\bq + ... = ... + {\bf B}\bu$ is satisfied, using the S-procedure.

Verifying dynamics in implicit form

Typically we write our differential equations in the form $\dot\bx = {\bf f}(\bx, \bu).$ But let us consider for a moment the case where the dynamics are given in the form $${\bf g}(\bx, \bu, \dot\bx ) = 0.$$ This form is strictly more general because I can always write ${\bf g}(\bx,\bu,\dot\bx) = f(\bx,\bu) - \dot\bx$, but importantly here I can also write the bottom rows of ${\bf g}$ as $\bM(\bq)\ddot\bq + \bC(\bq,\dot\bq)\dot\bq - \btau_g - \bB \bu$. This form can also represent differential algebraic equations (DAEs) which are more general than ODEs; $\bg$ could even include algebraic constraints like $s_i^2 + c^2 - 1$. Most importantly, for manipulators, ${\bf g}$ can be polyonimal, even if ${\bf f}$ would have been rational. provides access to continuous-time dynamics in implicit form via the CalcImplicitTimeDerivativesResidual method.

Interestingly, we can check the Lyapunov conditions, $\dot{V}(\bx) \le 0$, directly on a system (with no inputs) in its implicit form, $\bg(\bx,\dot\bx)=0$. Simply define a new function $Q(\bx, \bz) = \frac{\partial V(\bx)}{\partial \bx} \bz.$ If we can show $Q(\bx, \bz) \le 0, \forall \bx,\bz \in \{ \bx, \bz | \bg(\bx,\bz) = 0 \}$ using SOS, then we have verified that $\dot{V}(\bx) \le 0$, albeit at the non-trivial cost of adding indeterminates $\bz$ and an additional S-procedure.

There are a few things that do break this clean polynomial view of the world. Rotary springs, for instance, if modeled as $\tau = k (\theta_0 - \theta)$ will mean that $\theta$ appears alongside $\sin\theta$ and $\cos\theta$, and the relationship between $\theta$ and $\sin\theta$ is sadly not polynomial. Linear feedback from LQR actually looks like the linear spring, although writing the feedback as $u = -\bK \sin\theta$ is a viable alternative.

In practice, you can also Taylor approximate any smooth nonlinear function using polynomials. This can be an effective strategy in practice, because you can limit the degrees of the polynomial, and because in many cases it is possible to provide conservative bounds on the errors due to the approximation.

One final technique that is worth knowing about is a change of coordinates, often referred to as the stereographic projection, that provides a coordinate system in which we can replace $\sin$ and $\cos$ with polynomials:

The stereographic projection.
By projecting onto the line, and using similar triangles, we find that $p = \frac{\sin\theta}{1 + \cos\theta}.$ Solving for $\sin\theta$ and $\cos\theta$ reveals $$\cos\theta = \frac{1-p^2}{1+p^2}, \quad \sin\theta = \frac{2p}{1+p^2}, \quad \text{and} \quad \frac{\partial p}{\partial \theta} = \frac{1+p^2}{2},$$ where $\frac{\partial p}{\partial \theta}$ can be used in the chain rule to derive the dynamics $\dot{p}$. Although the equations are rational, they share the denominator $1+p^2$ and can be treated efficiently in mass-matrix form. Compared to the simple substitution of $s=\sin\theta$ and $c=\cos\theta$, this is a minimal representation (scalar to scalar, no $s^2+c^2=1$ required); unfortunately it does have a singularity at $\theta=\pi$, so likely cannot be used for global analysis.

Control design

Throughout this chapter, we've developed some very powerful tools for reasoning about stability of a closed-loop system (with the controller already specified). I hope you've been asking yourself -- can I use these tools to design better controllers? Of course the answer is "yes!". In this section, I'll discuss the control approaches that are the most direct consequences of the convex optimization approaches to Lyapunov functions. Another very natural idea is to use these tools in the content of a feedback motion planning algorithm, which is the subject of an upcoming chapter.

State feedback for linear systems

Let's re-examine the Lyapunov conditions for linear systems from Eq. \ref{eq:lyap}, but now add in a linear state feedback $\bu = \bK\bx$, resulting in the closed-loop dynamics $\dot{\bx} = (\bA + \bB\bK)\bx.$ One can show that the set of all stabilizing $\bK$ can be characterized by the following two matrix inequalities Boyd94: \[{\bf P} = {\bf P}^T \succ 0, \qquad {\bf P}(\bA + \bB\bK) + (\bA + \bB\bK)^T{\bf P} \prec 0. \] Unfortunately, these inequalities are bilinear in the decision variables, ${\bf P}$ and $\bK$, and therefore non-convex. It turns out that we can turn these into linear matrix inequalities (LMIs) through a simple change of coordinates, $\bQ = {\bf P}^{-1}, {\bf Y} = \bK{\bf P}^{-1}$, resulting in \[ \bQ = \bQ^T \succ 0, \qquad \bA\bQ + \bQ\bA^T + \bB{\bf Y} + {\bf Y}^T \bB^T \prec 0. \] Furthermore, given matrices $\bA$ and $\bB$, there exists a matrix $\bK$ such that $(\bA + \bB\bK)$ is stable if and only if there exist matrices $\bQ$ and ${\bf Y}$ which satisfy this (strict) linear matrix inequality.

The robust state-feedback design for polytopic uncertainty example from LMIsinSYSTEMSCONTROL.pdf (in Zotero) by Denis ARZELIER essentially combines this with the linear common-lyapunov function example.

Control design via alternations

For control design using convex optimization, we will lean heavily on the assumption of the dynamics being control-affine. Let me write it this time as: $$\dot{\bx} = {\bf f}(\bx) + \sum_{i=0}^{m-1} u_i {\bf f}_i(x).$$ As we have discussed, for mechanical systems this assumption is perfectly reasonable.

For linear optimal control, we found controllers of the form, $\bu = -\bK \bx.$ The natural generalization of this to polynomial analysis will be to look for controllers of the form $\bu = \pi(\bx)$, where $\pi(\bx)$ is a polynomial. (For mechanical systems like the pendulum above, we might make $\pi$ a polynomial in $s$ and $c$.)

If we apply this control to the Lyapunov conditions (for global analysis), we quickly see the problem. We have $$\dot{V}(\bx) = \pd{V}{\bx}\left[ {\bf f}(x) + \sum_{i=0}^{m-1} {\bf f}_i(\bx) \pi_i(\bx)\right],$$ and as a result if we are trying to search for the parameters of $V$ and $\pi$ simultaneously, the decision variables multiply and the problem will be non-convex.

One very natural strategy is to use alternations. The idea is simple, we will fix $\pi$ and optimize $V$, then fix $V$ and optimize $\pi$ and repeat until convergence. This approach has roots in the famous "DK iterations" for robust control (e.g. Lind94). It takes advantage of the structured convex optimization at each step.

For this approach, it is important to start with an initial feasible $V$ or $\pi$. For example, one can think of locally stabilizing a nonlinear system with LQR, and then searching for a linear control law (or even a higher-order polynomial control law) that achieves a larger basin of attraction. But note that once we move from global stability to region of attraction optimization, we now need to alternate between three sets of variables: $V(\bx), \pi(\bx), \lambda(\bx),$ where $\lambda(]bx)$ was the multiplier polynomial for the S-procedure. We took exactly this approach in Majumdar13. In that paper, we showed that alternations could increase the certified region of attraction for the Acrobot.

The alternation approach takes advantage of convex optimization in the inner loop, but it is still only a local approach to solving the nonconvex joint optimization. It is subject to local minima. The primary advantage is that, barring numerical issues, we expect recursive feasibility (once we have a controller and Lyapunov function that achieves stability, even with a small region of attraction, we will not lose it) and monotonic improvement on the objective. It is also possible to attempt to optimize these objectives more directly with other non-convex optimization procedures (like stochastic gradient descent, or sequential quadratic programming) (e.g. Shen20), but strict feasibility is harder to guarantee. Often times the Lyapunov conditions are just evaluated at samples, or along sample trajectories; we can still certify the performance using just a single SOS verification step the with controller fixed before deplying.

Control-Lyapunov Functions

Another core idea connecting Lyapunov analysis with control design is the "control-Lyapunov function". Given a system $\dot{\bx} = f(\bx, \bu)$, a control-Lyapunov function $V$ is a positive-definite function for which $$\forall \bx \neq 0, \exists \bu \quad \dot{V}(\bx,\bu) = \pd{V}{\bx} f(\bx, \bu) < 0 \quad \text{and} \quad \exists \bu \quad \dot{V}(0, \bu) = 0.$$ In words, for all $\bx$, there exists a control that would allow the $V$ to decrease. Once again, we adorn this basic property with extra conditions (e.g. radially unbounded, or satisfied over a control- invariant set) in order to construct global or regional stability statements. What is important to understand is that we can design control-Lyapunov functions without explicitly parameterizing the controller; often the control action is not even determined until execution time by finding a particular $\bu$ that goes downhill.

Our sums-of-squares toolkit is well-suited for addressing questions with the $\forall$ quantifier over indeterminates. Working with the $\exists$ quantifier is much more difficult; we don't have an S-procedure-like solution for it. Interestingly, there is one approach that we've discussed above that effectively turns the $\exists$ into a $\forall$ -- that is the outer approximation approach to region of attraction analysis.

In the outer-approximation, we produce a barrier certificate to find the set of states where the controller cannot go (for any $\bu$). Our barrier certificate now has the form $$\forall x, \forall u, \quad \dot{\mathcal{B}}(\bx, \bu) \le 0.$$ Certainly, $\mathcal{B}(\bx) \ge 0$ is now an outer-approximation of the true "backward-reachable set" (BRS) of the fixed point Henrion14+Korda13. Again Posa17 has some nice examples of this written directly in sums-of-squares form.

Now here is where it gets a little frustrating. Certainly, sublevel sets of $\mathcal{B}(\bx)$ are control-invariant (via the proper choice of $\bu$). But we do not (cannot) expect that the entire estimated region (the 0-superlevel set) can be rendered invariant. The estimated region is an outer approximation of the backward reachable set. it's more than that... \dot{B}=0 happens somewhere in the interior, and it's conservative due to the forall u. Majumdar13b gave a recipe for extracting a polynomial control law from the BRS; an inner approximation of this control law can be obtained via the original SOS region of attraction tools. This is unfortunatley suboptimal/conservative. Although we would like to certify the control-Lyapunov conditions directly in an inner approximation, the $\exists$ quantifier remains as an obstacle.

Approximate dynamic programming with SOS

We have already established the most important connection between the HJB conditions and the Lyapunov conditions: $$\dot{J}^*(\bx) = -\ell(\bx,\bu^*) \qquad \text{vs} \qquad \dot{V}(\bx) \preceq 0.$$ The HJB involves solving a complex PDE; by changing this to an inequality, we relax the problem and make it amenable to convex optimization.

Upper and lower bounds on cost-to-go

Asking for $\dot{V}(\bx) \preceq 0$ is sufficient for proving stability. But we can also use this idea to provide rigorous certificates as upper or lower bounds of the cost-to-go. Given a control-dynamical system $\dot{\bx} = f(\bx, \bu)$, and a fixed controller $\pi(\bx)$ we can find a function $V(\bx)$:

  • $\forall \bx, \dot{V}^\pi(\bx) \le -\ell(\bx,\pi(\bx))$ to provide an upper bound, or
  • $\forall \bx, \dot{V}^\pi(\bx) \ge -\ell(\bx,\pi(\bx))$ to provide an lower bound.
To see this, take the integral of both sides along any solution trajectory, $\bx(t), \bu(t)$. For the upper-bound we get \begin{gather*} \int_0^\infty \dot{V}^\pi(\bx) dt = V^\pi(\bx(\infty)) - V^\pi(\bx(0)) \le \int_0^\infty -\ell(\bx(t),\pi(\bx(t))) dt.\end{gather*} Assuming $V^\pi(\bx(\infty)) = 0$, we have $$V^\pi(\bx(0)) \ge \int_0^\infty \ell(\bx(t),\pi(\bx(t))) dt.$$

The upper bound is the one that we would want to use in a certification procedure -- it provides a guarantee that the total cost achieved by the system started in $\bx$ is less than $V(\bx)$. But it turns out that the lower bound is much better for control design. This is because we can write $$\forall \bx, \min_\bu [ \ell(\bx,\bu) + \pd{V}{\bx}f(\bx,\bu) ] \ge 0 \quad \equiv \quad \forall \bx, \forall \bu, \ell(\bx,\bu) + \pd{V}{\bx}f(\bx,\bu) \ge 0.$$ Therefore, without having to specify apriori a controller, if we can find a function $V(\bx)$ such that $\forall \bx, \forall \bu, \dot{V}(\bx, \bu) \ge -\ell(\bx,\bu)$, then we have a lower-bound on the optimal cost-to-go.

You should take a minute to convince yourself that, unfortunately, the same trick does not work for the upper-bound. Again, we would need $\exists$ as the quantifier on $\bu$ instead of $\forall$.

Sums-of-squares formulation

Combining a number of ideas we've seen already, this leads to a natural sums-of-squares formulation for optimal control: \begin{align*} \max_{{V}(\bx)} \quad & \int_{\bf X} V(\bx)d\bx, \\ \subjto \quad & \ell(\bx,\bu) + \pd{V}{\bx}f(\bx,\bu) & \text{ is SOS}, \\ & V(0) = 0.\end{align*} The SOS constraint enforces the lower bound, and the objective "pushes up" on the lower-bound by maximizing the integral over some compact region. Once again, we can then try to extract a control law by either using this lower bound as a control-Lyapunov function, or by extracting (and certifying) a polynomial controller.

Perhaps you noticed that this is a natural extension of the linear programming approach to dynamic programming. For systems with continuous state, the LP approach verified the inequality conditions only at sample points; here we verify them for all $\bx, \bu$. This is an important generalization: not so much because it can certify better (the lower bound is not a particularly valuable thing to certify), but because it can scale to dimensions were dense sampling is not viable. This provides something of a spectrum between the mesh-based value iteration and the very scalable LQR.

The biggest challenge to this approach, however, is not the number of dimensions, but the degree of the polynomial required to achieve a meaningful approximation. Remember that the optimal cost-to-go for even the min-time double integrator problem is non-smooth. Like in the pseudo-spectral methods that we will see in the context of trajectory optimization; choosing the right polynomial basis can make a huge difference for the numerical robustness of this approach.

SOS ADP for the cubic polynomial

SOS ADP for the pendulum swing-up problem

Alternative computational approaches

"Satisfiability modulo theories" (SMT)

Satisiability modulo theories (SMT). dReal is available in .

Mixed-integer programming (MIP) formulations

Continuation methods

Neural Lyapunov functions

Contraction metrics

Control contraction metrics
other topics/ideas: verifying neural network control relation to adversarial examples. make the point that sampling doesn't scale (even with polynomials). "almost lyapunov functions" will happen when we talk about robust control. hopfield enegy function. scaling (e.g. quadratic in high dimensions with just one strictly negative eigenvalue); would be very hard to verify with samples.


Valid Lyapunov Function for Global Stability

For the system \begin{align*} \dot x_1 &=-\frac{6x_1}{(1+x_1^2)^2}+2x_2, \\ \dot x_2 &=-\frac{2(x_1+x_2)}{(1+x_1^2)^2}, \end{align*} you are given the positive definite function $V(\bx) =\frac{x_1^2}{1 + x_1^2}+ x_2^2$ (plotted here) and told that, for this system, $\dot V(\bx)$ is negative definite over the entire space. Is $V(\bx)$ a valid Lyapunov function to prove global asymptotic stability of the origin for the system described by the equations above? Motivate your answer.

Invariant Sets and Regions of Attraction

You are given a dynamical system $\dot \bx = f(\bx)$, with $f$ continuous, which has a fixed point at the origin. Let $B_r$ be a ball of (finite) radius $r > 0$ centered at the origin: $B_r = \{ \bx : \| \bx \| \leq r \}$. Assume you found a continuously-differentiable scalar function $V(\bx)$ such that: $V(0) = 0$, $V(\bx) > 0$ for all $\bx \neq 0$ in $B_r$, and $\dot V(\bx) < 0$ for all $\bx \neq 0$ in $B_r$. Determine whether the following statements are true or false. Briefly justify your answer.

  1. $B_r$ is an invariant set for the given system, i.e.: if the initial state $\bx(0)$ lies in $B_r$, then $\bx(t)$ will belong to $B_r$ for all $t \geq 0$.
  2. $B_r$ is a subset of the ROA of the fixed point $\bx = 0$, i.e.: if $\bx(0)$ lies in $B_r$, then $\lim_{t \rightarrow \infty} \bx(t) = 0$.

Are Lyapunov Functions Unique?

If $V_1(\bx)$ and $V_2(\bx)$ are valid Lyapunov functions that prove global asymptotic stability of the origin, does $V_1(\bx)$ necessarily equal $V_2(\bx)$?

Proving Global Asymptotic Stability

Consider the system given by \begin{align*} \dot x_1 &= x_2 - x_1^3, \\ \dot x_2 &= - x_1 - x_2^3. \end{align*} Show that the Lyapunov function $V(\bx) = x_1^2 + x_2^2$ proves global asymptotic stability of the origin for this system.

Gradient Flow in Euclidean Space

We can use Lyapunov analysis to analyze the behavior of optimization algorithms like gradient descent. Consider an objective function $\ell:\mathbf{R}^n\rightarrow \mathbf{R}$, and we want to minimize $\ell(x)$. The gradient flow is a continuous-time analog of gradient descent and is defined as $\dot{x} = -\frac{\partial \ell}{\partial x}$.

  1. Show that the objective function $\ell(x)-\ell(x^*)$ is a Lyapunov function of the gradient flow where $x^*$ is a unique minimizer.
  2. We can use Lyapunov to argue that an optimization problem will converge to a global optimum, even if it is non-convex. Suppose that the Lyapunov function $\ell$, has negative definite $\dot{\ell}$. Show that the objective function $\ell$ has a unique minimizer at the origin.
  3. Consider the objective function in the figure below. Could this be a valid Lyapunov function for showing global asymptotic stability?
  4. Lyapunov function candidate for asymptotic stability.
  5. Consider the objective function $\ell(x) = x_1^4-8x_1^3+18x_1^2+x_2^2$ with $x\in\mathbf{R}^2$. Find the largest $\rho$ such that $\{x\mid \ell(x)<\rho \}$ is a valid region of attraction for the origin.

Control-Lyapunov Function for a Wheeled Robot

In this exercise, we examine the idea of a control-Lyapunov function to drive a wheeled robot, implementing the controller proposed in Aicardi95.

Similar to this previous example, we use a kinematic model of the robot. We represent with $z_1$ and $z_2$ its Cartesian position and with $z_3$ its orientation. The controls are the linear $u_1$ and angular $u_2$ velocities. The equations of motion read \begin{align*} \dot z_1 &= u_1 \cos z_3, \\ \dot z_2 &= u_1 \sin z_3, \\ \dot z_3 &= u_2.\end{align*} The goal is to design a feedback law $\pi(\bz)$ that drives the robot to the origin $\bz=0$ from any initial condition. As pointed out in Aicardi95, this problem becomes dramatically easier if we analyze it in polar coordinates. As depicted below, we let $x_1$ be the radial and $x_2$ the angular coordinate of the robot, and we define $x_3 = x_2 - z_3$. Analyzing the figure, basic kinematic considerations lead to \begin{align*} \dot x_1 &= u_1 \cos x_3, \\ \dot x_2 &= - \frac{u_1 \sin x_3}{x_1}, \\ \dot x_3 &= - \frac{u_1 \sin x_3}{x_1} - u_2.\end{align*}

Wheeled robot with Cartesian $\bz$ and polar $\bx$ coordinate system.
  1. For the candidate Lyapunov function $V(\bx) = V_1(x_1) + V_2(x_2, x_3)$, with $V_1(x_1) = \frac{1}{2} x_1^2$ and $V_2(x_2, x_3) = \frac{1}{2}(x_2^2 + x_3^2)$, compute the time derivatives $\dot V_1 (\bx, u_1)$ and $\dot V_2(\bx, \bu)$.
  2. Show that the choice \begin{align*} u_1 &= \pi_1(\bx) = - x_1 \cos x_3, \\ u_2 &= \pi_2(\bx) = x_3 + \frac{(x_2 + x_3) \cos x_3 \sin x_3}{x_3}, \end{align*} makes $\dot V_1 (\bx, \pi_1(\bx)) \leq 0$ and $\dot V_2 (\bx, \pi(\bx)) \leq 0$ for all $\bx$. (Technically speaking, $\pi_2(\bx)$ is not defined for $x_3=0$. In this case, we let $\pi_2(\bx)$ assume its limiting value $x_2 + 2 x_3$, ensuring continuity of the feedback law.)
  3. Explain why Lyapunov's direct method does not allow us to establish asymptotic stability of the closed-loop system.
  4. Substitute the control law $\bu = \pi (\bx)$ in the equations of motion, and derive the closed-loop dynamics $\dot \bx = f(\bx, \pi(\bx))$. Use LaSalle's theorem to show (global) asymptotic stability of the closed-loop system.
  5. In this python notebook we set up a simulation environment for you to try the controller we just derived. Type the control law from point (b) in the dedicated cell, and use the notebook plot to check your work.

Limitations of SOS Polynomials in Lyapunov Analysis

  1. Are there positive definite functions that are not representable as sums of squares?
  2. If a fixed point of our dynamical system does not admit a SOS Lyapunov function, what can we conclude about its stability?

ROA Estimation for the Time-Reversed Van der Pol Oscillator

In this exercise you will use SOS optimization to approximate the ROA of the time-reversed Van der Pol oscillator (a variation of the classical Van der Pol oscillator which evolves backwards in time). In this python notebook, you are asked to test the following SOS formulations.

  1. The one from the example above, augmented with a line search that maximizes the area of the ROA.
  2. A single-shot SOS program that can directly maximize the area of the ROA, without any line search.
  3. An improved version of the previous, where less SOS constraints are imposed in the optimization problem.


  1. Jean-Jacques E. Slotine and Weiping Li, "Applied Nonlinear Control", Prentice Hall , October, 1990.

  2. Hassan K. Khalil, "Nonlinear Systems", Prentice Hall , December, 2001.

  3. S. Boyd and L. El Ghaoui and E. Feron and V. Balakrishnan, "Linear Matrix Inequalities in System and Control Theory", SIAM , 1994.

  4. Pablo A. Parrilo, "Structured Semidefinite Programs and Semialgebraic Geometry Methods in Robustness and Optimization", PhD thesis, California Institute of Technology, May 18, 2000.

  5. Amir Ali Ahmadi and Miroslav Krstic and Pablo A. Parrilo, "A Globally Asymptotically Stable Polynomial Vector Field with no Polynomial Lyapunov Function", Proceedings of the Conference on Decision and Control , 2011.

  6. Shen Shen and Russ Tedrake, "Sampling Quotient-Ring Sum-of-Squares Programs for Scalable Verification of Nonlinear Systems", Proceedings of the 2020 59th IEEE Conference on Decision and Control (CDC) , 2020. [ link ]

  7. Russ Tedrake and Ian R. Manchester and Mark M. Tobenkin and John W. Roberts, "{LQR-Trees}: Feedback Motion Planning via Sums of Squares Verification", International Journal of Robotics Research, vol. 29, pp. 1038--1052, July, 2010. [ link ]

  8. Anirudha Majumdar, "Funnel Libraries for Real-Time Robust Feedback Motion Planning", PhD thesis, Massachusetts Institute of Technology, Jun, 2016. [ link ]

  9. Didier Henrion and Milan Korda, "Convex computation of the region of attraction of polynomial control systems", IEEE Transactions on Automatic Control, vol. 59, no. 2, pp. 297-312, 2014.

  10. Jean Bernard Lasserre, "Moments, Positive Polynomials and Their Applications", World Scientific , vol. 1, 2010.

  11. Michael Posa and Twan Koolen and Russ Tedrake, "Balancing and Step Recovery Capturability via Sums-of-Squares Optimization", Robotics: Science and Systems, 2017. [ link ]

  12. Gerald B Folland, "How to integrate a polynomial over a sphere", The American Mathematical Monthly, vol. 108, no. 5, pp. 446--448, 2001.

  13. D. Henrion and J.B. Lasserre and C. Savorgnan, "Approximate volume and integration for basic semialgebraic sets", SIAM Review, vol. 51, no. 4, pp. 722--743, 2009.

  14. Mark M. Tobenkin and Ian R. Manchester and Russ Tedrake, "Invariant Funnels around Trajectories using Sum-of-Squares Programming", Proceedings of the 18th IFAC World Congress, extended version available online: arXiv:1010.3013 [math.DS], 2011. [ link ]

  15. Anirudha Majumdar, "Robust Online Motion Planning with Reachable Sets", , May, 2013. [ link ]

  16. Charles W. Wampler and Andrew J. Sommese, "Numerical algebraic geometry and algebraic kinematics", Acta Numerica, vol. 20, pp. 469-567, 2011.

  17. A.J. Sommese and C.W. Wampler, "The Numerical solution of systems of polynomials arising in engineering and science", World Scientific Pub Co Inc , 2005.

  18. R. Lind and G.J. Balas and A. Packard, "Evaluating {D-K} iteration for control design", American Control Conference, 1994 , vol. 3, pp. 2792 - 2797 vol.3, 29 June-1 July, 1994.

  19. Anirudha Majumdar and Amir Ali Ahmadi and Russ Tedrake, "Control Design along Trajectories with Sums of Squares Programming", Proceedings of the 2013 IEEE International Conference on Robotics and Automation (ICRA), pp. 4054-4061, 2013. [ link ]

  20. Milan Korda and Didier Henrion and Colin N Jones, "Controller design and region of attraction estimation for nonlinear dynamical systems", The 19th World Congress of the International Federation of Automatic Control (IFAC), 2014.

  21. Anirudha Majumdar and Ram Vasudevan and Mark M. Tobenkin and Russ Tedrake, "Convex Optimization of Nonlinear Feedback Controllers via Occupation Measures", Proceedings of Robotics: Science and Systems (RSS) , 2013. [ link ]

  22. Michele Aicardi and Giuseppe Casalino and Antonio Bicchi and Aldo Balestrino, "Closed loop steering of unicycle like vehicles via Lyapunov techniques", IEEE Robotics \& Automation Magazine, vol. 2, no. 1, pp. 27--35, 1995.

Previous Chapter Table of contents Next Chapter