For decades, the "gold standard" for ensuring the safety of autonomous systems has been Hamilton-Jacobi (HJ) reachability analysis. It provides a rigorous mathematical framework to determine the set of all states from which a robot can be guaranteed to avoid an obstacle or reach a goal, regardless of worst-case disturbances. However, this method has long been haunted by the "curse of dimensionality." Because traditional solvers rely on grid-based discretization of the state space, their computational requirements grow exponentially. In practice, this has limited formal safety guarantees to simple systems with fewer than six dimensions—leaving complex robots like quadrupeds, rockets, and multi-agent swarms in a "safety-critical" limbo.
A new research direction is breaking this bottleneck by replacing rigid grids with neural networks. By approximating the solutions to the underlying partial differential equations (PDEs) using deep learning, researchers are scaling formal safety verification to systems with 10, 20, and even 40 dimensions.
From Grids to Neural Value Functions
The core of HJ reachability is the value function, a map that assigns a score to every possible state of a system. If the score is positive, the system is safe; if it is negative, a collision is inevitable. Traditionally, calculating this function required solving the Hamilton-Jacobi-Isaacs (HJI) equation across a dense mesh.
Researchers like Somil Bansal and Claire Tomlin have pioneered the shift toward Neural Reachability. Instead of a grid, a neural network is trained to satisfy the HJI equation. This works because the value functions in these safety games are often Lipschitz continuous, a property that neural networks are exceptionally good at approximating. In Generating Formal Safety Assurances for High-Dimensional Reachability, the authors demonstrate that this "DeepReach" approach can identify safe operating regions for landing rockets and multi-vehicle avoidance—tasks that would require more memory than any modern supercomputer could provide if solved with grids.
Amortizing Safety with Predictive Filters
Once a safety value function is learned, it can be used as a "safety filter." This filter sits between a high-performance (but potentially unsafe) controller—like a neural network trained via reinforcement learning—and the robot's motors. If the controller suggests an action that would push the robot into a "danger zone" (a negative value in our learned function), the filter overrides it with a minimal intervention to keep the system safe.
The challenge is that evaluating safety in real-time is computationally expensive. Alexandre Didier and Melanie Zeilinger address this in Approximate predictive control barrier function for discrete-time systems by approximating the entire optimization process of a safety filter with a single neural network evaluation. This "amortized inference" allows for 1900x speedups, enabling high-frequency safety checks that can keep up with the fast dynamics of drones and legged robots.
Bridging Learning and Control
A recurring theme in this field is the integration of these neural safety certificates with existing control frameworks like Model Predictive Control (MPC) and Reinforcement Learning (RL).
- MPC Integration: Hao Wang and Somil Bansal show in Safe and Performant Deployment of Autonomous Systems how a learned safety value function can serve as a "terminal constraint" in an MPC loop. This ensures that even if a robot is being driven by a "hallucinating" Large Language Model, it remains within a provably safe boundary.
- RL Integration: In Safety and Liveness Guarantees through Reach-Avoid Reinforcement Learning, Kai–Chieh Hsu and Jaime Fernández Fisac reformulate reach-avoid objectives into Bellman-style equations. This allows RL agents to learn safety and goal-reaching simultaneously, scaling rigorous analysis to 6D systems and beyond.
Verifying the "Black Box"
If we are using a neural network to guarantee safety, how do we know the neural network itself is correct? This is the central tension of the field. To bridge the gap between "learned approximation" and "formal guarantee," researchers are developing hybrid verification techniques.
Albert Lin and Somil Bansal proposed a framework in Verification of neural reachable tubes via scenario optimization and conformal prediction that uses statistical sampling to provide probabilistic safety bounds. By treating the neural network's errors as a distribution, they can calculate an "outlier-adjusted" safety buffer. This ensures that even if the neural PDE solver is slightly off, the robot remains safe with a high, quantifiable probability.
Scaling to the Extreme
The most recent work in this area is pushing the boundaries of what was previously thought possible. Le Qiu and Zeyuan Feng have demonstrated that by using MPC to generate supervised labels for neural networks, reachability analysis can scale to 40-dimensional systems (Bridging Model Predictive Control and Deep Learning for Scalable Reachability Analysis).
Similarly, Justin Wang and Haimin Hu introduced MAGICS, an adversarial RL framework that provides safety filters for a 36-degree-of-freedom quadruped (MAGICS). These results suggest that the "curse of dimensionality" is no longer an insurmountable wall, but a hurdle that can be cleared by combining the rigor of control theory with the representational power of deep learning.