When designing controllers for complex, safety-critical systems—like walking robots, autonomous vehicles, or powered prostheses—we face a fundamental tension between foresight and stability.
Modern optimization-based control, such as Model Predictive Control (MPC), excels at foresight. By simulating a system's future behavior over a finite time horizon, an MPC controller can plan smooth trajectories, optimize for energy efficiency, and respect physical boundaries like torque or velocity limits. However, because this planning horizon is finite, MPC does not inherently guarantee long-term stability. A controller might find a path that looks optimal for the next ten steps, only to leave the system in an unrecoverable state at step eleven.
Traditionally, control engineers solve this by carefully tuning cost functions, adding heavy terminal penalties, or extending the planning horizon. This is a "soft" approach: it encourages stability by making instability expensive. But tuning these weights is notoriously tedious and fragile under unexpected disturbances.
Lyapunov-constrained predictive control solves this problem by turning stability into a "hard" geometric boundary within the optimization loop. Instead of hoping that a cost function will guide a system to safety, it mathematically forbids the controller from ever choosing an unstable action.
The Core Mechanism: Enforcing Energy Decay
To understand why this approach works, imagine a marble rolling inside a bowl. No matter where you release the marble, gravity and friction will eventually bring it to rest at the lowest point—the equilibrium.
In control theory, we generalize this intuition using a Lyapunov function, $V(x)$. You can think of $V(x)$ as a mathematical representation of the system's total "energy" at a given state $x$. If the system is exactly where we want it to be (the equilibrium), its energy is zero ($V(x) = 0$). Anywhere else, its energy is positive ($V(x) > 0$).
If we can guarantee that this energy strictly decreases with every step the system takes, the system will inevitably slide down the energy valley and stabilize at the equilibrium. Mathematically, we want to enforce a decrease condition at each time step $k$:
$$V(x_{k+1}) - V(x_k) \le -\alpha(x_k)$$
where $\alpha(x_k)$ is a positive function representing the minimum required rate of decay.
In a Control Lyapunov Function (CLF) framework, we look for a control input $u$ that satisfies this inequality. By embedding this exact inequality directly into an optimization problem (like a Quadratic Program or an MPC loop) as a hard constraint, we fundamentally alter the solver's search space. The optimization algorithm is free to maximize performance, cut power consumption, or seek out informative data, but it is strictly confined to a subset of control inputs that are mathematically guaranteed to drive tracking errors toward zero.
Unifying Foresight with Instantaneous Guarantees
Pure Lyapunov control is often myopic; it calculates the best action for the immediate next step without looking further ahead, which can lead to aggressive, high-gain control inputs. Conversely, standard MPC looks far ahead but lacks instantaneous stability guarantees over short horizons.
Integrating CLF constraints directly into predictive control creates a "best of both worlds" paradigm. As shown by Ruben Grandia et al. in Nonlinear Model Predictive Control of Robotic Systems with Control Lyapunov Functions, placing stability-guaranteeing constraints directly into a nonlinear MPC optimization eliminates the need for the tedious manual tuning of terminal costs or weights. The controller can plan complex maneuvers over a short horizon because stability is handled explicitly by the constraint, rather than implicitly by a long planning window.
This explicit prioritization of stability is crucial for highly dynamic or aggressive maneuvers. For instance, when a quadrotor transports a cable-suspended payload that might suddenly go slack, Luis F. Recalde et al. demonstrate in ES-HPC-MPC: Exponentially Stable Hybrid Perception Constrained MPC for Quadrotor with Suspended Payloads that embedding exponentially stabilizing CLFs as hard constraints ensures tracking errors converge to zero at a guaranteed rate, regardless of other competing optimization objectives. Similarly, for legged robots navigating unpredictable terrain, Maria Vittoria Minniti et al. show in Adaptive CLF-MPC With Application to Quadrupedal Robots that hard CLF inequalities force the optimization solver to select inputs that maintain a stable trajectory, even when the robot encounters massive, unmodeled payload changes.
Robustness and Filtering Under Uncertainty
Real-world systems rarely match their theoretical models perfectly, and they must constantly reject external disturbances. To maintain stability under uncertainty, the Lyapunov constraint must adapt.
One strategy is to decouple trajectory planning from low-level stabilization using a multi-rate architecture. In Multi-Rate Planning and Control of Uncertain Nonlinear Systems: Model Predictive Control and Control Lyapunov Functions, Andrew J. Taylor et al. utilize a high-level MPC planner to generate reference trajectories, while a faster, low-level CLF controller ensures robust tracking. The mathematical properties of the low-level CLF are used to derive strict bounds for the high-level planner, ensuring that the entire system satisfies physical constraints even when buffeted by disturbances.
When system dynamics are learned from data rather than derived from physics, we can treat stability probabilistically. Armin Lederer et al. demonstrate in Parameter Optimization for Learning-based Control of Control-Affine Systems how combining Lyapunov stability theory with scenario optimization allows a system to guarantee tracking performance with high probability, avoiding the overly aggressive, high-gain control actions typical of traditional worst-case theoretical bounds. This stable-by-design philosophy is also explored by Anna Scampicchio et al. in Stable and robust LQR design via scenario approach. By embedding Lyapunov requirements as Linear Matrix Inequalities (LMIs) into a finite-horizon Linear Quadratic Regulator (LQR), the authors map out a fundamental control trade-off: accepting a minor reduction in immediate performance efficiency to buy absolute, guaranteed long-term stability.
[ Arbitrary Control Input ]
│
▼
┌───────────────────────────────────┐
│ Predictive Stability Filter │ ◄─── Enforces V(x_k+1) < V(x_k)
└───────────────────────────────────┘
│
▼
[ Safe & Asymptotically Stable Input ]
Alternatively, instead of redesigning the core predictive controller, engineers can deploy a predictive stability filter as an external safety layer. As developed by Andrea Zanelli et al. in Predictive stability filters for nonlinear dynamical systems affected by disturbances, this filter intercepts control inputs and modifies them only as much as necessary to satisfy a Lyapunov function decrease constraint. This architecture ensures robust asymptotic stability under disturbances and simplifies implementation, as the decrease can be implicitly enforced within the optimization problem without needing to explicitly compute complex nonlinear functions for the entire system.
Real-World Applications
By providing formal guarantees without sacrificing real-time performance, Lyapunov-constrained optimization has become a cornerstone for complex robotic applications.
- Assistive and Rehabilitation Robotics: In prosthetics and orthoses, movements must be smooth, stable, and highly responsive to human intent. As synthesized by Samer Mohammed et al. in the Special Issue on Assistive and Rehabilitation Robotics, Lyapunov-based optimization ensures precise tracking of human gait phases while maintaining energy efficiency. A practical hardware validation of this is detailed in Demonstration of locomotion with the powered prosthesis AMPRO utilizing online optimization-based control, where embedding CLFs into a real-time quadratic programming framework boosted tracking accuracy by 23% and cut power consumption by 25% compared to traditional methods.
- Human-Robot Collaboration: When robots physically interact with humans, safety cannot be compromised. Loris Roveda et al. in Q-Learning-Based Model Predictive Variable Impedance Control for Physical Human-Robot Collaboration (Extended Abstract)* integrate Lyapunov constraints into a variable impedance MPC framework, allowing a robot to adapt its stiffness and damping in real-time to optimize collaboration while guaranteeing physical interaction safety.
- Autonomous Vehicles and Swarms: For vehicles operating in challenging environments, Daxiong Ji et al. show in Three-Dimensional Dynamic Positioning Using a Novel Lyapunov-Based Model Predictive Control for Small Autonomous Surface/Underwater Vehicles that Lyapunov-based MPC provides a rigorous mathematical foundation for recursive feasibility and closed-loop stability in complex underwater conditions.
The framework also enables multi-objective missions. For instance, Olga Napolitano et al. demonstrate in Information-Aware Lyapunov-Based MPC in a Feedback-Feedforward Control Strategy for Autonomous Robots that a Lyapunov stability constraint can be successfully paired with an information-seeking feedforward component, forcing an autonomous robot to converge toward its goal while actively steering through paths that maximize sensor data quality. At the swarm level, Anna Lukina demonstrates in Adaptive Optimization Framework for Control of Multi-Agent Systems that dynamically adjusting the planning horizon of an MPC can implicitly construct a Lyapunov function, analytically guaranteeing that a multi-agent swarm will safely converge from random initial positions into an optimal, energy-efficient formation.
Key Foundational Work
Demonstration of locomotion with the powered prosthesis AMPRO utilizing online optimization-based control (2015) — contributed 1 concept elements to this direction.
Special Issue on Assistive and Rehabilitation Robotics (2017) — contributed 1 concept elements to this direction.