ShardNet embeds safety into neural network architecture for 100% constraint enforcement

ShardNet, developed by Long Kiu Chung and Shreyas Kousik, tackles a fundamental challenge in deploying neural network controllers on safety-critical systems: how to enforce strict, non-convex constraints without compromising performance. Traditional methods treat safety as a metric to optimize alongside other objectives, often resulting in trade-offs or training failures. ShardNet flips this by embedding safety into the neural network's structure. It uses a differentiable projection layer parameterized by a classification network to strictly enforce unions of polyhedral constraints and learned value function level sets — a first for non-convex constraints in safe-by-construction neural architectures.

On double integrator benchmarks from the literature, ShardNet policies maintained 100% safety on verified sets while achieving significantly lower objective loss compared to existing formal methods. The paper also introduces a technique to verify and train rectified linear unit (ReLU) networks as value functions, which had not been possible before. This allows safe sets to be expanded more than 3 times larger than prior verification approaches, paving the way for safer autonomous vehicles, drones, and robotics where rigorous constraint satisfaction is non-negotiable.