Successive CLFs

Chaining control Lyapunov functions to reach a goal through clutter.

Back to research

Driving a system to a goal through clutter rarely admits a single polynomial control Lyapunov function. Successive CLFs split the job: each certificate only has to bring the system into the region of a lower level, and the last one brings it home.

level 0 (goal) to level 10 not certified wall

A toy model, built so the mechanism is visible: a point robot with bounded speed in a room with one wall, and quadratic CLFs whose regions are the ellipses. Each region avoids the wall and contains the centre of the next-lower one. Click a free spot to launch a robot. The chart plots its active Lyapunov function, which decreases inside each level and restarts lower at every hand-over. In the method itself, the CLFs are polynomials for nonlinear systems, synthesized with SOS programming.

The idea

A control Lyapunov function \(V\) certifies that a system \(\dot x = f(x, u)\) can be driven to a goal: \(V\) is positive away from the goal, and wherever it applies there is an input with

\[\nabla V(x) \cdot f(x, u) \;\le\; -\alpha\, V(x).\]

On a cluttered domain, the region where one polynomial \(V\) can meet this condition while steering clear of obstacles is small.

Successive CLFs use a sequence \(V_0, V_1, \dots, V_N\) with regions \(\mathcal{D}_k = \{x : V_k(x) \le 1\}\). The base certificate \(V_0\) is an ordinary CLF around the goal. For \(k \ge 1\), \(V_k\) only needs to decrease on \(\mathcal{D}_k\) outside the lower regions, and its closed-loop trajectories must enter \(\mathcal{D}_0 \cup \dots \cup \mathcal{D}_{k-1}\) without becoming unsafe. Every level is a separate SOS program, so the hierarchy grows outward from the goal one certificate at a time.

At run time, the controller selects the lowest level whose region contains the current state and follows that level’s CLF until a lower level takes over. Levels only go down, so each level is used at most once and there are at most \(N\) switches; per-level transit times add up to a bound on the time to reach the goal (Wajid, 2026).

What to look for in the demo

  • A robot starting at the far end of the room passes through every level in order; its trail changes colour at each hand-over.
  • In the chart, each tooth of the sawtooth is one level’s \(V_k\) running down. It never rises within a level.
  • The hatched corners are states no level certifies. Adding levels, or richer polynomial certificates, shrinks them.

References

2026

  1. PHD
    Successive Control Certificates for Safe Autonomy
    Rameez Wajid
    University of Colorado at Boulder, 2026