BarrierNet: A Safety-Guaranteed Layer for Neural Networks

Abstract

This paper introduces differentiable higher-order control barrier functions (CBF) that are end-toend trainable together with learning systems. CBFs are usually overly conservative, while guaranteeing safety. Here, we address their conservativeness by softening their definitions using environmental dependencies without loosing safety guarantees, and embed them into differentiable quadratic programs. These novel safety layers, termed a BarrierNet, can be used in conjunction with any neural network-based controller, and can be trained by gradient descent. BarrierNet allows the safety constraints of a neural controller be adaptable to changing environments. We evaluate them on a series of control problems such as traffic merging and robot navigations in 2D and 3D space, and demonstrate their effectiveness compared to state-of-the-art approaches.

Differentiable Control Barrier Functions for Vision-based End-to-End Autonomous Driving

“…Differentiable optimization-based safety frameworks. Recent advances in differentiable optimization methods show promise for safety guaranteed neural network controllers [5,36,6,53]. In [5], a differentiable quadratic program (QP) layer, called OptNet, was introduced.…”

“…They are offline methods, unable to adapt to environment changes (such as varying size of the unsafe sets) [31]. By comparison, the BarrierNet in [53] incorporates differentiable CBFs into neural network controllers. In this work, we address the challenges of using BarrierNet to achieve vision-based endto-end autonomous driving.…”

“…In this subsection, we briefly introduce BarrierNet [53] that incorporates differentiable CBFs, and propose solutions for some of the challenges that arise in autonomous driving with BarrierNet. Differentiable CBFs are motivated by the fact that the traditional CBFs can easily make the system overly conservative.…”

“…Prior work also introduced differentiable CBFs as novel CBF formulations whose parameters are trainable [51], [34]. They have been incorporated into differentiable QPs [5] which can in turn be combined with learning systems [53].…”

“…In CBFs, a system tends to stay far away from the unsafe set boundary, and thus may deviate largely from a desired trajectory. A BarrierNet [53] is constructed by incorporating dCBFs into a differentiable QP. In a BarrierNet, all QP parameters, including those originating in CBFs, are trainable.…”