Safe-by-Design NN controllers avoid costly QP filters

Traditional safety filters for neural network controllers rely on control barrier functions (CBFs) with hand-tuned, fixed parameters. These decoupled designs often become overly conservative and require solving an online quadratic program (QP) at runtime, creating a computational bottleneck for real-time applications like autonomous driving or robotics. Yang Zhao and colleagues address this by jointly learning the controller and CBF parameters as neural networks, ensuring that the control policy inherently satisfies affine safety constraints by construction — no online QP needed.

To further boost scalability, the team introduces a lightweight projection architecture that enforces constraints without enumerating the entire set. This design keeps model size small while maintaining provable safety guarantees. Extensive simulations across multiple benchmark environments show the approach reliably hits safety targets with significantly lower compute than prior methods. The work bridges the gap between formal safety verification and practical deployment, enabling faster, safer autonomous systems without sacrificing performance.