Safe Control Under Input Limits with Neural Control Barrier Functions

We scale safety filter synthesis to high-dimensional systems by posing it as training a neural control barrier function to satisfy CBF conditions. We synthesize a safety filter for a 10D quadrotor–pendulum system with less than 2 hours of training; the filter triggers orders of magnitude less often than model predictive control (MPC) and prevents the pendulum from falling while the quadrotor tracks a trajectory.