⚡ electrical engineering

Safety-Aware Performance Boosting for Constrained Nonlinear Systems

This paper proposes a safety-aware control architecture that integrates a performance-boosting controller with a scheduled Predictive Safety Filter to guarantee stability and safety while strictly expanding the set of achievable trajectories beyond conventional methods, enabling complex behaviors like transient detours in constrained nonlinear systems.

Original authors: Danilo Saccani, Haoming Shen, Luca Furieri, Giancarlo Ferrari-Trecate

Published 2026-03-23

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Danilo Saccani, Haoming Shen, Luca Furieri, Giancarlo Ferrari-Trecate

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are teaching a robot to walk through a crowded, dangerous room. The robot has two main goals:

Stay Safe: It must not bump into walls or fall over (constraints).
Be Efficient: It needs to get to its destination quickly and smoothly, even if that means taking a weird detour to avoid a moving person.

For a long time, engineers had to choose between these two. If they made the robot super careful (safety-first), it would move like a turtle, unable to take shortcuts. If they let it be fast (performance-first), it might crash.

This paper introduces a clever new "control architecture" that lets the robot be both safe and smart. Here is how it works, broken down into simple concepts:

The Two-Team System

The authors split the robot's brain into two distinct roles:

1. The "Safety Filter" (The Strict Parent)
Think of this as a strict parent or a safety guard. Its only job is to make sure the robot never breaks the rules.

It has a Lyapunov Certificate, which is like a "safety battery."
In traditional systems, this battery must drain a little bit every single second. If the battery gets too low, the system is safe.
The Problem: Because the battery must drain constantly, the robot is trapped in a shrinking circle of safety. It can't take a risk, even a temporary one, to go around an obstacle. It's like a child who is told they must walk in a straight line toward home; if a puddle is in the way, they can't step around it because stepping sideways would "waste" their safety budget.

2. The "Performance Booster" (The Creative Driver)
This is the smart, learning part of the brain. It figures out the best, fastest way to get to the goal.

It suggests moves, like "Let's take a detour!" or "Let's speed up!"
However, it doesn't have direct control. It has to ask the Safety Filter for permission.

The Magic Trick: The "Scheduling" Mechanism

The breakthrough in this paper is how these two talk to each other.

In old systems, the Safety Filter was rigid: "I will only let you move if your safety battery drains by exactly 10% every second."

In this new system, the Performance Booster gets to schedule the drain rate.

The Metaphor: Imagine the Safety Battery has a "drain valve."
Usually, the valve is set to a steady, slow drip (safe but slow).
But, if the Performance Booster sees a moving obstacle (like a person walking by), it can say, "Hey, I need to take a quick detour. Can we open the valve a little wider for just a few seconds?"
The Safety Filter agrees, but with a catch: "Okay, I'll let you drain the battery faster right now, but only because you are promising to drain it extra fast later to make up for it."

Why This Matters: The "Detour"

The paper proves mathematically that this flexibility allows the robot to do things that were previously impossible.

The Scenario: A robot needs to balance a pole (an inverted pendulum) while a wall moves toward it.
The Old Way: The robot sees the wall. To stay safe, it must keep its "safety battery" draining steadily. It calculates that to avoid the wall, it needs to swing the pole wildly to the side. But that swing would drain the battery too fast right now. The Safety Filter says "No," and the robot gets stuck or crashes.
The New Way: The Performance Booster says, "I'll swing the pole hard to the side to dodge the wall!" The Safety Filter checks the schedule, sees that the robot is in a "transient" phase, and says, "Okay, I'll allow the battery to drain faster for a moment. Just make sure you stabilize it once the wall is gone."
The Result: The robot successfully dodges the wall, then immediately stabilizes itself, ending up safe and on track.

How They Taught the Robot (Training)

Teaching a robot to do this is hard because the Safety Filter is a complex math puzzle (an optimization problem) that is difficult to "teach" a computer to learn from.

The authors used a clever trick:

Instead of trying to teach the robot how to solve the Safety Filter's math puzzle directly (which is like trying to teach a student how to solve a test while they are taking the test), they treated the Safety Filter as a black box.
They used a "Critic" (a teacher) to watch the robot. If the robot suggests a move and the Safety Filter accepts it, the Critic gives a good grade. If the Safety Filter rejects it, the Critic gives a bad grade.
The robot learns from these grades without ever needing to understand the complex math inside the Safety Filter.

The Bottom Line

This paper solves a major problem in robotics: How do we make machines that are both safe and capable of complex, dynamic tasks?

By letting the "Performance" part of the brain temporarily negotiate with the "Safety" part, the system can take calculated risks (like a detour) that were previously forbidden, as long as it guarantees safety in the long run. It's like giving a driver a map that says, "You can drive fast and take shortcuts, as long as you promise to slow down and stop at the red light eventually."

1. Problem Formulation

The paper addresses the challenge of controlling nonlinear, constrained systems where the goal is to optimize performance (execute complex tasks) while strictly guaranteeing safety (state/input constraints) and stability (closed-loop convergence).

The Conflict: Traditional Model Predictive Control (MPC) ensures safety via hard constraints but can be conservative when optimizing performance. Conversely, Learning-based approaches (e.g., Reinforcement Learning) offer high performance but often lack formal safety and stability guarantees.
The Limitation of Existing Safety Filters: Predictive Safety Filters (PSFs) are used to modify unsafe control inputs to satisfy constraints. However, standard PSFs enforce a fixed, monotonic Lyapunov decrease rate. This forces the system to remain within shrinking Lyapunov level sets, preventing complex behaviors like "safe detours" around obstacles that might temporarily increase the Lyapunov function value.
The Goal: Develop a control architecture that decouples safety, stability, and performance, allowing the system to temporarily relax stability constraints to achieve better performance, while mathematically guaranteeing that the system eventually returns to a stable state.

2. Methodology

The authors propose a control architecture (Figure 1) consisting of two main components: a Performance-Boosting (PB) Controller and a Scheduled Predictive Safety Filter (PSF).

A. The Scheduled Predictive Safety Filter (PSF)

The PSF acts as a pre-stabilizing base controller. Unlike standard PSFs that enforce a fixed Lyapunov decrease rate ( $\rho$ ), this paper introduces a scheduling mechanism:

Variable Decrease Rate: The required Lyapunov decrease rate, $\rho_t$ , is no longer fixed. It is dynamically adjusted based on the magnitude of the performance input ( $u_L$ ) proposed by the PB controller.
Tightening Schedule ( $\psi$ ): A function $\psi$ $ψ$ maps the norm of the performance input ( $\|u_{L,t}\|$ $∥ u_{L, t} ∥$ ) to the scheduling parameter $\rho_t$ $ρ_{t}$ .
- Transient Phase: When the performance input is large (indicating a complex maneuver or detour), $\rho_t$ is allowed to increase (even $\rho_t > 1$ ), effectively relaxing the Lyapunov decrease constraint. This permits the system to temporarily move "uphill" in the Lyapunov landscape to navigate obstacles.
- Stabilization Phase: As the performance input diminishes ( $\|u_{L,t}\| \to 0$ ), $\rho_t$ automatically tightens to a fixed value $\bar{\rho} < 1$ , ensuring the system eventually converges to the equilibrium.
Optimization: At each step, the PSF solves a constrained optimization problem to find the safe input $u$ closest to $u_L$ that satisfies state/input constraints and the scheduled Lyapunov decrease constraint.

B. The Performance-Boosting (PB) Controller

Parameterization: The PB controller ( $M_\theta$ ) is parameterized as a causal operator using a Magnitude-and-Direction (MAD) policy.
Stability by Design: The architecture ensures that the output $u_L$ belongs to the $\ell_2$ space (finite energy) by construction. This is achieved by using a Linear Recurrent Unit (LRU) for the magnitude component, which guarantees $\ell_2$ -stability for any parameters.
Role: It generates the "desired" control input for performance optimization and provides the signal to the scheduler to adjust the PSF's strictness.

C. Training Procedure

Black-Box Approach: Directly differentiating through the PSF's optimization problem is difficult due to non-smoothness when constraints become active.
Actor-Critic RL: The authors use an off-policy actor-critic method (e.g., DDPG).
- The Actor learns the policy $u_L = \mu_\theta(\zeta)$ .
- The Critic estimates the value of $u_L$ .
- Crucially, the gradient update does not require differentiating through the PSF solver. Instead, it relies on the gradient of the Q-function with respect to $u_L$ , treating the PSF-plant interconnection as a deterministic black box.

3. Key Contributions & Theoretical Results

(i) Stability by Design

The paper proves Theorem 1: If the PB controller generates an $\ell_2$ -stable input ( $u_L \in \ell_2$ ), the closed-loop system is guaranteed to be $\ell_2$ -stable and satisfy all constraints.

Mechanism: Even though $\rho_t$ can be $\ge 1$ transiently, the fact that $u_L \to 0$ ensures that $\rho_t$ eventually settles to $\bar{\rho} < 1$ , guaranteeing asymptotic stability.

(ii) Trajectory-Set Expansion

The paper proves Theorem 2: The set of achievable safe trajectories using the Scheduled PSF strictly contains the set of trajectories achievable by a Fixed-Rate PSF.

Significance: This mathematically proves that the proposed architecture can execute behaviors (like transient detours) that are provably impossible for standard PSFs, which are confined to shrinking Lyapunov level sets.

(iii) Differentiable-Free Training

The proposed training procedure bypasses the need to differentiate through the non-smooth PSF optimization map, making the learning of complex policies feasible and stable.

4. Experimental Results

The method was validated on a constrained inverted pendulum task with a moving obstacle.

Task: Stabilize the pendulum at the upright equilibrium ( $\theta=0$ ) while avoiding a moving obstacle that blocks the direct path.
Comparison:
- Baseline (Fixed-Rate PSF): Failed the task. The monotonic decrease of the Lyapunov function forced the pendulum to stay within shrinking level sets, making it impossible to detour around the obstacle without violating the stability constraint.
- Proposed (Scheduled PSF): Successfully completed the task. The scheduler allowed the Lyapunov certificate ( $J^*_t$ ) to increase transiently (by setting $\rho_t > \bar{\rho}$ ), enabling the pendulum to swing away from the equilibrium to avoid the obstacle. Once the obstacle was cleared, the input magnitude dropped, $\rho_t$ tightened, and the system stabilized.
Visualization: Figure 3 shows the Lyapunov certificate for the proposed method rising temporarily (blue line) before converging, whereas the baseline (orange line) is forced to decrease monotonically, leading to failure.

5. Significance

This work bridges the gap between formal safety guarantees and high-performance control in nonlinear systems.

Overcoming Conservatism: It solves the fundamental limitation of fixed-rate safety filters that prevent complex, non-monotonic behaviors.
Theoretical Rigor: It provides formal proofs that relaxing stability constraints transiently does not compromise long-term stability, provided the performance input is $\ell_2$ -bounded.
Practical Applicability: By decoupling the safety filter from the learning process via a black-box approach, it enables the use of powerful RL techniques for safety-critical systems without requiring complex differentiable optimization layers.

In summary, the paper introduces a "safety-aware" boosting mechanism that allows autonomous systems to take calculated, temporary risks (detours) to achieve complex goals, while mathematically guaranteeing they will return to a safe, stable state.