RAYEN researchers report faster neural networks that enforce operating constraints

Researchers report that RAYEN, a framework for neural networks, guarantees that network outputs or internal variables satisfy specified convex constraints during training and testing. The framework avoids repeated corrective calculations and supports several constraint types used in robotics. The paper reports 8 ms of additional processing time to impose 1K quadratic constraints on a 1K-dimensional variable. In constrained trajectory optimization tests, RAYEN ran 20 to 7468 times faster than state-of-the-art algorithms while maintaining constraint satisfaction and near-optimal cost. Researchers also tested a learned quadruped locomotion policy in simulation and real-world experiments, validating its compliance with actuator constraints. The reported guarantees concern the supported convex constraints.