[Paper Review] Safe Control with Learned Certificates: A Survey of Neural Lyapunov, Barrier, and Contraction methods

Charles Dawson, Sicun Gao+1 more|arXiv (Cornell University)|Feb 23, 2022

Fault Detection and Control Systems19 citations

TL;DR

This survey introduces neural certificate methods—learning Lyapunov, barrier, and contraction functions via neural networks—to provide verifiable safety and stability guarantees for learned controllers in nonlinear robotics systems. By jointly training control policies and differentiable certificates, the framework enables safe, stable, and verifiable control without relying on hand-designed functions or restrictive system assumptions, offering a scalable alternative to traditional methods like sum-of-squares or PDE-based approaches.

ABSTRACT

Learning-enabled control systems have demonstrated impressive empirical performance on challenging control problems in robotics, but this performance comes at the cost of reduced transparency and lack of guarantees on the safety or stability of the learned controllers. In recent years, new techniques have emerged to provide these guarantees by learning certificates alongside control policies -- these certificates provide concise, data-driven proofs that guarantee the safety and stability of the learned control system. These methods not only allow the user to verify the safety of a learned controller but also provide supervision during training, allowing safety and stability requirements to influence the training process itself. In this paper, we provide a comprehensive survey of this rapidly developing field of certificate learning. We hope that this paper will serve as an accessible introduction to the theory and practice of certificate learning, both to those who wish to apply these tools to practical robotics problems and to those who wish to dive more deeply into the theory of learning for control.

Motivation & Objective

• To provide a comprehensive, accessible survey of neural certificate methods for safe and stable control in nonlinear dynamical systems.
• To bridge the gap between formal control theory and modern learning-based control by unifying Lyapunov, barrier, and contraction function learning under a neural network framework.
• To address the lack of guarantees in deep reinforcement learning by enabling self-supervised training with verifiable certificates.
• To identify open challenges and future research directions in heterogeneous multi-agent, distributed, and networked control systems.
• To serve as a foundational reference for practitioners and researchers applying or advancing certificate learning in robotics and control.

Proposed method

• The framework uses neural networks to represent control certificates—Lyapunov, barrier, and contraction functions—enabling end-to-end differentiable training with learned controllers.
• Certificates are trained using a self-supervised signal derived from the certificate conditions (e.g., negative definite derivative for Lyapunov functions), eliminating the need for explicit reward shaping.
• The method supports non-polynomial dynamics and high-dimensional systems, overcoming limitations of sum-of-squares and PDE-based approaches.
• It integrates with reinforcement learning by allowing the certificate to guide policy optimization through implicit supervision, ensuring stability and safety during training.
• The approach enables verifiable correctness: a trained neural certificate provides a mathematical proof of system stability or safety for the corresponding controller.
• Compositional verification and graph neural networks are proposed for extending the framework to heterogeneous and distributed multi-agent systems.

Experimental results

Research questions

• RQ1 Can neural networks effectively learn Lyapunov, barrier, and contraction functions that provide verifiable stability and safety guarantees for nonlinear systems?
• RQ2 How can certificate learning be integrated with reinforcement learning to provide both performance and safety guarantees?
• RQ3 What are the limitations of existing methods like sum-of-squares and PDE-based approaches, and how do neural certificates overcome them?
• RQ4 Can neural certificates be scaled to large-scale, high-dimensional, or multi-agent systems with heterogeneous dynamics?
• RQ5 What are the key challenges in applying neural certificates to distributed, networked, or fault-tolerant control systems?

Key findings

• Neural certificates enable the synthesis of verifiable safety and stability guarantees for nonlinear systems without requiring hand-designed functional forms or polynomial dynamics.
• The framework supports self-supervised training, where the certificate conditions themselves provide the supervisory signal, reducing reliance on external reward functions.
• The method has been successfully applied to complex tasks such as stable walking under uncertainty, quadrotor flight in turbulent environments, and decentralized control of over 1,000 agents.
• Neural certificates outperform traditional methods in scalability and generality, avoiding the computational intractability of solving high-dimensional PDEs or the restrictive assumptions of sum-of-squares programming.
• The approach provides a practical path toward trustworthy, learning-based control in real-world robotics applications, with strong theoretical grounding and empirical validation.
• Future work is needed to extend the framework to heterogeneous multi-agent systems, distributed control with delays, and fault-tolerant architectures.