In robotics, safety is often a binary choice between rigor and scale. On one hand, classical control theory offers "gold standard" tools like Hamilton-Jacobi (HJ) reachability analysis and Lyapunov stability proofs. These provide formal guarantees that a robot will never enter an unsafe state. However, as Lukas Brunke, Angela P. Schoellig, and their colleagues note in Safe Learning in Robotics, these methods suffer from the "curse of dimensionality." They typically fail when a system has more than five or six degrees of freedom because they rely on dense computational grids that grow exponentially with the number of state variables.
On the other hand, Deep Reinforcement Learning (RL) scales to high-dimensional systems like humanoids or fighter jets but offers only "best-effort" safety. A robot might work 99% of the time, but that 1% failure is often catastrophic.
A new research direction is bridging this gap by replacing rigid computational grids with Neural Certificates. By using neural networks to represent safety boundaries and stability functions, researchers are now providing formal guarantees for systems with 10, 20, or even more dimensions.
The Geometry of a Safety Bubble
To understand why this works, imagine a robot's "safe set" as a bubble in space. If the robot is inside the bubble, there exists a control strategy to keep it from ever hitting an obstacle. If it drifts outside, no matter what it does, a collision is inevitable.
Calculating the exact shape of this bubble involves solving the Hamilton-Jacobi-Isaacs (HJI) partial differential equation. Traditionally, this requires carving the world into a fine grid. If you double the dimensions of your robot (e.g., moving from a 2D point to a 4D car), the number of grid points doesn't just double—it squares. This is why John Lygeros and Kostas Margellos originally framed these as differential games, where the controller "fights" against the worst-case disturbances to stay inside the bubble.
Replacing Grids with Neural Networks
The core insight of recent work is that we don't need a grid to define a bubble; we can use a neural network as a continuous function approximator. David Fridovich-Keil, Vicenç Rúbies Royo, and Claire Tomlin demonstrated that classification-based neural networks can learn these safety boundaries in high-dimensional spaces where grids are impossible.
Once the safety boundary is represented by a neural network, the problem shifts from "how do we store the grid?" to "how do we ensure the network is accurate?" Researchers have developed several strategies to handle this:
- Decomposition: Mo Chen and Sylvia Lee Herbert showed in Decomposition of Reachable Sets that high-dimensional systems can be broken into lower-dimensional subsystems. By computing safety for these smaller pieces and overlapping them, they can verify a 10D quadrotor in a fraction of the time.
- Adversarial Training: In MADR, Ryan Teoh and Sander Tonkens use Model Predictive Control (MPC) to guide an adversarial training process. This allows the system to discover the "worst-case" scenarios that might pop the safety bubble, scaling formal reachability to 20-dimensional systems.
- Policy Evaluation: Oswin So and Chuchu Fan introduced Policy Neural Control Barrier Functions (PNCBFs), which reframe safety as a policy evaluation problem. This bypasses the numerical instability of solving PDEs directly, enabling safety guarantees for a 16D F-16 jet model.
Proving the Certificate
A neural network is just a "guess" at a safety boundary until it is verified. To turn these networks into formal certificates, researchers use tools from optimization and formal methods.
Hongkai Dai and Benoit Landry pioneered the use of Mixed-Integer Linear Programming (MILP) to verify that a neural network controller satisfies Lyapunov stability conditions. Similarly, Yujie Yang and Changliu Liu developed a scalable synthesis framework that uses MILP to formally verify neural value functions for HJ reachability. If the solver finds no violations of the safety conditions, the neural network becomes a mathematically rigorous proof of safety.
Online Adaptation and Resilience
The ultimate goal is for these certificates to adapt as the world changes. Static safety buffers are often too conservative—a drone might fly too slowly because it’s "afraid" of a wind gust that isn't there.
Sander Tonkens and Jorge Cortés addressed this with SPACE2TIME, which recasts spatial disturbances as temporal ones. This allows a robot to use an offline-learned safety certificate and "stretch" or "shrink" it online based on the actual disturbances it feels. By combining this with confidence-based predictions from Somil Bansal, robots can now maintain formal safety even when interacting with unpredictable humans or changing environments.
By moving from discrete grids to continuous neural certificates, the field is finally enabling robots that are both highly capable and provably safe.