Near-Optimal Constrained Feedback Control of Nonlinear Systems via Approximate HJB and Control Barrier Functions
A new two-stage method decouples performance from safety, allowing real-time constraint changes without retraining.
A team of researchers has published a new paper on arXiv outlining a significant advance in AI-driven control systems for robotics and autonomous vehicles. The framework, developed by Milad Alipour Shahraki and Laurent Lessard, tackles the critical challenge of ensuring both high performance and strict safety in complex, nonlinear systems like drones, robotic arms, and spacecraft. Their method cleverly separates the problem into two parts: an offline "brain" that learns an efficient control strategy, and an online "safety guard" that ensures all actions are safe.
The core innovation is the decoupling of performance optimization from constraint enforcement. First, an approximate value function is computed offline by solving the Hamilton-Jacobi-Bellman (HJB) equation, which defines a near-optimal control policy for an unconstrained system. Then, during real-time operation, a lightweight quadratic program is solved. This program minimally adjusts the optimal policy's commands to strictly satisfy safety constraints, which are mathematically defined using Control Barrier Functions (CBFs). This architecture means safety rules—like "stay 1 meter from this wall"—can be added, removed, or changed on the fly without the computationally expensive need to retrain the entire control model from scratch.
The researchers validated their approach on two benchmark problems: a simple 2-state linear hovercraft and a complex, nonlinear 9-state spacecraft attitude control system. The results demonstrated that their controller achieved performance very close to computationally heavy, offline optimal control benchmarks, while significantly outperforming existing methods based on Control Lyapunov Functions. This represents a practical step toward more adaptable and trustworthy autonomous systems that can operate reliably in dynamic, real-world environments.
- Two-stage architecture decouples performance (offline HJB solution) from safety (online CBF-based quadratic program).
- Enables safety constraints to be modified in real-time without retraining the core control policy.
- Validated on a complex 9-state nonlinear spacecraft model, showing near-optimal performance vs. benchmarks.
Why It Matters
Enables safer, more adaptable robots and autonomous systems that can update safety rules instantly for dynamic environments.