BarrierFormer: Transformer-Guided Predictive Barrier Enforcement for Safe Robot Control

This paper introduces BarrierFormer, a barrier-supervised transformer framework designed to address the myopic nature of standard control barrier functions (CBFs) and the computational cost of model predictive control (MPC). By using a causal transformer to encode observation-action history and autoregressively generate predictive rollouts, the system replaces the need for a known dynamic model during online optimization. A barrier critic evaluates constraint violations along these rollouts, while a safety teacher provides direct supervision for learning safe policies. Evaluations on linear and nonlinear 2D and 3D dynamical systems demonstrate that BarrierFormer outperforms existing reinforcement learning, diffusion-based, MPC-based, and transformer-based approaches in both safety rate and inference latency, enabling real-time model-free predictive safety enforcement.