Learning Safe-by-Design Neural Network Controllers

Technical Summary: Learning Safe-by-Design Neural Network Controllers

Problem Statement
Neural networks (NNs) are increasingly employed as feedback control policies to handle complex, high-dimensional nonlinear dynamics. However, purely learned controllers often lack formal safety guarantees, potentially generating actions that violate critical constraints such as collision avoidance or actuator limits. A common mitigation strategy involves appending a safety filter, typically based on Control Barrier Functions (CBFs), to a pre-trained controller. In standard frameworks, this filter solves an online Quadratic Program (QP) to enforce safety constraints defined by a CBF candidate $h(x)$ and an extended class-$\mathcal{K}$ function $\alpha(\cdot)$.

This decoupled design presents two primary limitations:

1. Conservatism and Suboptimality: The $\alpha(\cdot)$ function is often fixed (e.g., $\alpha(h) = \omega h$) with manually tuned parameters. Poor tuning leads to overly conservative behavior, degrading control performance, while aggressive tuning risks safety violations.
2. Computational Overhead: Solving a QP at runtime introduces computational latency and reliability concerns, hindering real-time deployment, especially for systems with high-dimensional state spaces or numerous constraints.

While recent works like BarrierNet and CAffNet attempt to integrate safety into the learning process, they either retain the expensive online QP layer or suffer from scalability issues due to the exponential complexity of projecting outputs onto feasible sets defined by many constraints.

Methodology
The authors propose a "Safe-by-Design" framework that jointly learns a neural network controller and neural-network-parameterized CBF conditions, embedding safety constraints directly into the controller architecture to eliminate the need for online QP solvers.

1. CAffNet-Lite Architecture:
   The core contribution is CAffNet-Lite, a lightweight extension of the existing CAffNet architecture. It integrates a constraint-affine layer with a trainable projection module to satisfy state-dependent affine constraints $A(x)u \leq b(x)$ by construction.

   • Lightweight Constraint Decomposition: Unlike CAffNet, which enumerates all possible combinations of active constraints (leading to exponential complexity $O(2^{n_c})$), CAffNet-Lite employs a reduced decomposition strategy. It considers only sub-constraints with the minimum order ($k=1$) or the maximum order ($k=\min(n_c, m)$). This reduces the worst-case complexity to polynomial $O(n_c^m)$, significantly improving scalability for systems with many constraints.
   • Projection Mechanism: The network outputs an unconstrained control signal $f_\theta(x)$. If this signal violates constraints, a projection layer computes a feasible candidate set by projecting onto the sub-constraint hyperplanes. The final output is the unconstrained signal (if feasible) or the closest feasible projection, ensuring hard constraint satisfaction without requiring full-rank or linearly independent constraints.

2. Joint Learning of CBF Parameters:
   Instead of fixing the decay function $\alpha(\cdot)$, the framework replaces it with a trainable neural network parameterization $\alpha_\vartheta(x, h(x))$. This allows the system to learn an "optimal-decay" property that adapts to the controller's behavior and system dynamics. The safety constraints are formulated as:
   $$A(x)u \leq b_\vartheta(x)$$
   where $b_\vartheta(x)$ incorporates the learned $\alpha_\vartheta$.

3. Theoretical Guarantees:
   Since the projection layer involves an $\arg\min$ operation that may result in discontinuous control signals, the authors establish safety in the Filippov sense. They prove that the resulting piecewise continuous control law guarantees strong forward invariance of the safe set, ensuring that if the system starts within the safe set, it remains there for all future time.

Key Contributions

1. CAffNet-Lite: Introduction of a lightweight projection architecture that reduces computational complexity from exponential to polynomial with respect to the number of constraints while preserving hard affine constraint satisfaction and universal approximation properties.
2. Joint Optimization: A framework that jointly trains a neural network controller and CBF parameters (specifically the decay function), enabling the satisfaction of safety constraints by construction without online QP solvers.
3. Theoretical Proof: Proof that the proposed controller satisfies strong forward invariance of the safe set in the Filippov sense, addressing the discontinuity introduced by the projection layer.

Experimental Results
The framework was evaluated on two simulation scenarios:

1. Single Integrator with Collision Avoidance:

   • Setup: A 2D integrator system avoiding three convex polyhedral obstacles.
   • Findings: CAffNet and CAffNet-Lite achieved the lowest mean cost (closest to the nominal controller) while strictly satisfying all safety constraints (0% violation). In contrast, standard NNs and HardNet exhibited constraint violations, and QP-based methods with fixed parameters were either too conservative (failing to reach the goal) or required higher computational time.
   • Efficiency: CAffNet-Lite matched the performance of CAffNet but with reduced testing time, demonstrating the efficiency of the lightweight decomposition.

2. Fixed-Wing Aircraft Geofencing:

   • Setup: A 7-state, 3-input aircraft model enforcing geofence constraints.
   • Findings: Both CAffNet and CAffNet-Lite successfully learned safe velocity commands that avoided geofences.
   • Efficiency: CAffNet-Lite demonstrated significantly lower training and testing times compared to the original CAffNet, confirming its superior scalability for higher-dimensional systems. While CAffNet converged slightly faster in terms of cost due to more projection candidates, CAffNet-Lite offered a better trade-off for real-time applicability.

Significance and Claims
The paper claims that by embedding affine safety constraints directly into the neural controller and learning the CBF parameters jointly, the proposed method overcomes the conservatism of fixed-parameter filters and the computational burden of online QP solvers. The authors emphasize that CAffNet-Lite provides a scalable, real-time deployable solution that guarantees hard safety constraints by construction. The work positions itself as a step toward "Safe-by-Design" control, where safety is not an afterthought or a runtime filter, but an intrinsic property of the learned policy. The authors modestly note that future work will explore extending the framework to broader classes of constraints and evaluating performance on real-world systems.