The future of autonomous technology hinges on one critical factor: unwavering safety. Groundbreaking research from institutions like the University of Waterloo and MIT CSAIL is paving the way by teaching AI systems to not just operate complex machines, but to mathematically prove their own stability and reliability. This fusion of deep learning with formal verification promises to transform how we build and trust everything from self-driving cars to power grids.
In the rapidly evolving landscape of artificial intelligence, the promise of autonomous systems—from industrial robots to self-driving vehicles and sophisticated power grids—is immense. Yet, a fundamental challenge persists: ensuring these dynamic machines function safely and predictably. When systems operate, they risk spiraling out of control unless meticulously designed with robust mathematical proof of stability. This crucial proof often relies on what’s known as a Lyapunov function, an “energy measure” indicating whether a system will safely settle into equilibrium. Historically, finding these functions, especially for systems influenced by randomness or non-linear forces, has been notoriously difficult.
The University of Waterloo’s Breakthrough: Learning Stability from Physics
A research team led by Dr. Jun Liu, Professor of Applied Mathematics and Canada Research Chair in Hybrid Systems and Control at the University of Waterloo, has developed an innovative method to automate this process. Their framework combines physics-informed neural networks—AI systems that learn directly from the laws of physics—with rigorous mathematical verification. This allows the AI to both learn and formally prove the stability of complex systems, a task traditionally performed manually by humans over months or years.
“Any time you’re dealing with a dynamic system—something that changes over time, such as an autonomous vehicle or a power grid—you can model it mathematically,” Liu explained. “Finding a function that guarantees stability, however, is often a notoriously difficult task.”
Teaching AI the Language of Stability
The core of Liu’s team’s innovation lies in teaching AI to discover Lyapunov functions. They reformulated two critical control theory equations, the Lyapunov and Zubov equations (types of partial differential equations), so that a neural network could learn them directly. If the network successfully satisfies these equations, it has effectively found a mathematical proof that the system is stable.
The Zubov equation extends this capability by mapping not only stability but also the full “safe zone” of operation, known as the domain of attraction. This crucial region precisely defines the conditions under which a machine can initiate operation and reliably return to safety without losing control. Once these conditions are learned, the network automatically generates a Lyapunov function, providing engineers with both a proof of stability and a visual map of where that stability holds.
The Crucial Step: Verifying AI’s Work with Formal Logic
While neural networks are powerful, their approximations sometimes fall short of rigorous proof. To counteract this, Liu’s team integrated a formal verification step. They employed an SMT solver (satisfiability modulo theories), a mathematical reasoning system, to meticulously check if the neural network’s proposed solution truly satisfied the strict inequalities required for stability. This two-step process—learning followed by proof—is what guarantees the model’s reliability.
The neural network first proposes a candidate Lyapunov function. Then, the SMT solver verifies two conditions: that the function remains positive everywhere except at equilibrium, and that its rate of change is consistently negative as the system evolves. If both conditions are met, the function is mathematically certified as valid. This hybrid approach synergizes human-designed mathematics with AI’s adaptive learning, combining simple, trusted local models with a neural network that can navigate broader, more complex stability regions where traditional methods struggle.
MIT CSAIL’s Contribution: Scaling Verification for Neural Network Controllers
In parallel, researchers at MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL) have also made significant strides in ensuring the stability of neural network-controlled robots. While neural networks have enabled more adaptive and efficient machines, their inherent complexity has made rigorous safety verification a significant hurdle, particularly when applying Lyapunov conditions to complex systems.
The MIT CSAIL team, including Lujie Yang and Russ Tedrake, developed new techniques that efficiently search for and verify Lyapunov functions, providing robust stability guarantees. Their innovative approach involves generating “cheaper counterexamples”—such as adversarial data from sensors—to optimize the robotic system. This process helps machines learn to handle challenging edge cases, enabling them to operate safely in a wider array of conditions. They also introduced a novel verification formulation utilizing a scalable neural network verifier, α,β-crown, to offer rigorous worst-case scenario guarantees beyond these counterexamples.
Lujie Yang, an MIT electrical engineering and computer science PhD student, highlighted the impact: “Our work bridges the gap between that level of performance from neural network controllers and the safety guarantees needed to deploy more complex neural network controllers in the real world.” This research was presented at the 2024 International Conference on Machine Learning, as reported by MIT News.
Performance Gains and Scalability
Both research groups have demonstrated superior performance compared to conventional methods. The Waterloo team’s AI-driven approach consistently outperformed traditional polynomial-based sum-of-squares techniques, which are limited to smooth, bowl-shaped energy surfaces. Neural networks, trained on the Lyapunov and Zubov equations, can adapt to the complex, non-linear stability regions common in real-world systems. For instance, in tests on the Van der Pol oscillator—a classic model for non-linear dynamics—the neural network accurately captured a significantly larger stability region.
To further enhance scalability, Liu’s team extended their method to systems with up to 20 interacting parts. They achieved this by verifying smaller subsystems individually and then combining the results, a compositional strategy that allows for efficient verification in large, interconnected networks like smart grids or fleets of autonomous drones.
The MIT CSAIL team also showed impressive results in simulations, guiding a quadrotor drone to a stable hover position using limited sensor data, and enabling stable operation for an inverted pendulum and a path-tracking vehicle under a wider range of conditions than previously possible for neural network verification methods.
Building Trust: The Future of Verifiable AI
The concept of using one AI system to verify another might seem counterintuitive, but Dr. Liu affirms its logic: “Neural networks can learn mathematical proofs of safety, while logic-based AI systems verify that those proofs are correct. Both are tasks humans used to perform manually.” This approach doesn’t seek to replace human engineers but rather to automate computationally intensive analysis, freeing human experts to focus on higher-level decision-making and innovation.
This groundbreaking work aligns with broader efforts to promote responsible AI through initiatives such as the TRuST Scholarly Network. By merging deep learning with formal mathematics, engineers can now design AI systems that are not only intelligent but also provably safe. This paradigm shift will redefine how we develop and certify advanced technologies across industries, from drones and self-driving cars to power grids and robotic surgical systems.
The vision is clear: in the near future, AI systems will not only control physical systems but also provide their own mathematical proof of safety before deployment. This potent combination of physics-based training and logical verification promises an era of AI controllers that regulators, engineers, and the public can fully trust. This research marks a pivotal step towards dependable machine learning, where systems that learn can also confidently guarantee their reliability, with findings published in the esteemed Journal Automatica.
