⚡ electrical engineering

Constrained Performance Boosting Control for Nonlinear Systems

This paper introduces an ADMM-based training framework for stable neural controllers that systematically enforces state and input constraints during performance boosting by alternating between gradient-descent updates and trajectory projection, achieving fewer constraint violations than soft-penalty baselines while maintaining stability-by-design guarantees.

Original authors: Gianluca Giacomelli, Danilo Saccani, Siep Weiland, Giancarlo Ferrari-Trecate, Valentina Breschi

Published 2026-08-05

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Gianluca Giacomelli, Danilo Saccani, Siep Weiland, Giancarlo Ferrari-Trecate, Valentina Breschi

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 trying to teach a robot to dance. You want it to move with incredible flair and speed, but you also have a strict rule: it must never step on the toes of the audience or trip over the stage lights. In the world of robotics and engineering, this is the eternal struggle between performance (how fast and well a machine does its job) and safety (making sure it doesn't break anything or hurt anyone).

For a long time, engineers have used "neural networks"—computer programs inspired by the human brain—to teach robots complex moves. These networks are like super-flexible dancers; they can learn almost any routine. However, there's a catch: because they are so flexible, it's hard to guarantee they won't accidentally step on a toe. Usually, engineers try to fix this by adding "penalties" to the training, like telling the robot, "If you step on a toe, you lose points." But this is like teaching a dancer by just shouting "Ouch!" when they make a mistake; it often leads to a robot that is either too scared to move (boring) or still steps on toes because it's trying too hard to win points.

The big question is: Can we teach a robot to be a star performer and a safety expert at the same time, without making the training process a mess? This is exactly what a team of researchers from Eindhoven University of Technology and EPFL in Switzerland set out to solve. They developed a new way to train these "neural" controllers that keeps the robot safe by design, while still letting it learn to dance perfectly.

The Problem with "Ouch!" Training

In the past, when researchers wanted to train a neural network to control a system (like a robot or a chemical plant) while respecting limits (like speed or position), they used a method called Performance Boosting (PB). Think of PB as a coach who tells the robot, "You are already safe because of how your body is built, now let's make you faster." The coach adjusts the robot's internal "weights" (its brain settings) to improve its moves.

However, the old PB method had a blind spot. It didn't know how to handle hard limits, like "Do not go faster than 0.5 meters per second." To fix this, previous attempts simply added a "soft penalty" to the training score. If the robot broke the rule, it got a big "ouch" (a penalty score). The problem with this approach is that it's a guessing game. If the "ouch" is too weak, the robot ignores the rules. If it's too strong, the robot becomes timid and moves slowly, ruining the performance. It's like trying to teach a child to ride a bike by either letting them crash or tying them to a wall; neither helps them learn to ride well.

The New Solution: ADMM-PB

The authors of this paper propose a clever new training routine called ADMM-PB. Instead of just shouting "Ouch!" when the robot breaks a rule, they use a mathematical strategy called the Alternating Direction Method of Multipliers (ADMM).

To understand how this works, imagine a dance rehearsal with two distinct roles:

The Dancer (The Controller): This is the robot's brain. Its only job is to try to make the dance look amazing and fast. It doesn't worry about the rules; it just focuses on the performance.
The Safety Coach (The Projection Step): This is a separate step that happens after the dancer tries a move. The coach looks at the dancer's moves and says, "Okay, that move was great, but you stepped on a toe. Let's gently nudge your foot back to a safe spot."

In the ADMM-PB method, these two roles take turns. First, the "Dancer" updates its brain to improve the performance. Then, the "Safety Coach" steps in and mathematically projects the robot's path onto the "safe zone," fixing any violations without changing the robot's actual brain structure. They repeat this back-and-forth process over and over.

This is a game-changer because it separates the two goals. The robot's brain remains in a "safe-by-design" shape (meaning it will never become unstable or dangerous, no matter what). The safety coach handles the rules. This means the robot doesn't have to guess how hard to be scared; it just learns to be safe while being fast.

What the Experiments Showed

The researchers tested this new method on two different scenarios to see if it actually worked.

The Robot Runner:
First, they simulated a point-mass robot trying to reach a target on a 2D plane while avoiding a circular obstacle. They compared their new ADMM-PB method against the old "soft penalty" method.

The Result: The old method was a rollercoaster. Depending on how hard they set the "ouch" penalty, the robot either crashed into the obstacle or moved too slowly. The new ADMM-PB method was much more consistent. It produced smoother training and, crucially, resulted in far fewer rule violations. The robot didn't just avoid the obstacle; it did so without becoming a timid turtle. The researchers found that while the new method was slightly more conservative (a bit more cautious), it was much better at balancing speed and safety than the old guessing-game approach.

The Three-Tank System:
Next, they tested it on a more complex system: three tanks of liquid connected by pipes, where the goal was to control the flow without overflowing or running dry.

The Result: Here, they started with a robot that had already been trained using the old "soft penalty" method. They then used ADMM-PB as a "refinement" step. Think of it as taking a good dancer and giving them a final polish. The new method improved the robot's ability to respect the limits (specifically, how fast the valve could change) by about 20%, without making the robot any slower or less stable. It proved that you can take an existing controller and make it safer without rebuilding it from scratch.

The Bottom Line

This paper doesn't claim to have solved every problem in the universe. The researchers are careful to note that while their method guarantees the robot stays stable (won't go crazy), it doesn't guarantee that the robot will never break a rule in every single possible future scenario, especially if the world is very chaotic. However, in their simulations, the new method consistently outperformed the old "penalty" approach.

The key takeaway is that by using this "Dancer and Coach" split (ADMM-PB), engineers can train neural controllers that are both high-performing and respectful of safety limits, without the frustrating trial-and-error of tuning penalty weights. It's a step toward robots that are not just smart, but also reliably safe partners in our world.

Technical Summary: Constrained Performance Boosting Control for Nonlinear Systems

Problem Statement
The paper addresses the challenge of designing high-performing controllers for stable nonlinear systems subject to state and input constraints. While Neural Network (NN) controllers have demonstrated impressive empirical performance in approximating complex feedback policies, they often lack formal guarantees regarding closed-loop stability and constraint satisfaction. Existing approaches to ensure stability, such as Internal Model Control (IMC) and nonlinear Youla parameterizations, allow for "stability-by-design" but typically treat constraints as secondary, often relying on ad-hoc barrier-inspired penalty terms in the loss function. These penalty-based methods require delicate tuning and do not systematically enforce constraints, potentially leading to either excessive constraint violations or overly conservative behavior. Conversely, online constraint enforcement methods (e.g., Model Predictive Control, safety filters) introduce significant computational complexity and architectural layers. The core problem is to develop an offline training procedure that enhances performance while strictly preserving the stability-by-design architecture and systematically handling convex state and input constraints without modifying the controller structure.

Methodology: ADMM-PB
The authors propose ADMM-PB (Alternating Direction Method of Multipliers for Performance Boosting), a learning procedure that extends the unconstrained Performance Boosting (PB) framework. The methodology relies on the following key components:

Stable-by-Design Parametrization: The controller is parametrized as a stable neural operator $\mathcal{M}(\theta)$ (e.g., a Recurrent Equilibrium Network) within the IMC architecture. This ensures that for any parameter value $\theta$ , the resulting closed-loop system is $\ell_p$ -stable, independent of the training algorithm's convergence.
Problem Reformulation: The finite-horizon Nonlinear Optimal Control (NOC) problem is reformulated by introducing auxiliary (copy) variables for the state and input trajectories ( $X_p, U_p$ ). This decouples the parameter optimization from the constraint enforcement.
ADMM Splitting: The training process alternates between two steps:
- Parameter Update ( $\theta$ -step): A gradient-descent-based update of the controller parameters. This step minimizes the performance loss plus a quadratic penalty term that measures the mismatch between the rollout trajectories generated by $\theta$ and the auxiliary copy variables. Crucially, this step retains the exact structure of the original unconstrained PB training, requiring no modification to the controller architecture.
- Projection Step ( $X_p, U_p$ -step): A feasibility step where the auxiliary trajectories are projected onto the convex sets of admissible states and inputs. This is a closed-form Euclidean projection that enforces hard constraints.
Dual Variable Update: Lagrange multipliers are updated to drive the primal residuals (the difference between rollout and projected trajectories) toward zero.

Key Contributions

Systematic Constraint Handling: The paper introduces a systematic mechanism to handle convex state and input constraints during the offline training of neural controllers, moving beyond the reliance on soft penalty terms.
Preservation of Stability Guarantees: The proposed method ensures that the closed-loop stability guarantee, inherent to the chosen stable parametrization, is preserved at every iteration of the ADMM routine. This guarantee is independent of the convergence of the ADMM algorithm or the termination criterion.
Decoupled Objectives: The approach cleanly separates performance-oriented parameter training, projection-based constraint handling, and stability preservation.
Practical Implementation: The authors provide practical guidelines for tuning hyperparameters (step sizes, termination criteria) and analyze the computational complexity, noting that the additional cost is dominated by the projection step, which is linear in the state/input dimensions for simple convex sets.

Results
The effectiveness of ADMM-PB is validated on two benchmark examples:

Point-Mass Robot: A robot navigating a 2D plane with a circular obstacle and velocity constraints. Compared to a baseline using barrier-inspired soft penalties, ADMM-PB achieved significantly lower constraint violations. While the penalty-based baseline could achieve faster convergence to the origin with low penalty weights, it resulted in high constraint violations. ADMM-PB offered a better trade-off, producing more consistent trajectories across scenarios with fewer violations, albeit with slightly more conservative behavior. The method also demonstrated smoother training dynamics compared to the penalty-based approach.
Three-Tank System: A classical control example used to demonstrate ADMM-PB as a refinement layer. Starting from a PB controller trained with soft penalties, ADMM-PB was used to further reduce constraint violations (specifically on input bounds and input variations) without degrading control performance. This confirmed the method's ability to refine existing solutions without altering the underlying architecture.

Significance and Claims
The paper claims that ADMM-PB provides a natural and effective way to incorporate constraints into the training of stable neural controllers without compromising the "stability-by-design" property. The authors emphasize that while the method improves constraint satisfaction and performance trade-offs compared to penalty-based methods, it does not guarantee global optimality or absolute constraint satisfaction in the testing phase if the ADMM iterations do not fully converge or if disturbances have unbounded support. The stability guarantee is strictly tied to the controller parametrization, not the training convergence. The work positions ADMM-PB as a practical offline training tool that bridges the gap between flexible neural control and rigorous constraint handling, offering a structured alternative to online optimization or complex certificate-based designs. The authors acknowledge that theoretical convergence guarantees for this specific nonconvex, rollout-based ADMM setting remain an open area for future research.

The Problem with "Ouch!" Training

The New Solution: ADMM-PB

What the Experiments Showed

The Bottom Line

More like this