A Control Barrier Function-Constrained Model Predictive Control Framework for Safe Reinforcement Learning

A team from NYU Tandon School of Engineering has proposed a new framework called Probabilistic Ensembles with CBF-constrained Trajectory Sampling (PECTS) to tackle a core challenge in reinforcement learning: ensuring safety when an AI agent operates in an unpredictable environment. The framework cleverly merges two established control theory concepts. First, it uses Model Predictive Control (MPC) to plan actions by predicting future system states. Second, it integrates Control Barrier Functions (CBFs)—mathematical conditions that define a "safe set" of states—directly into the MPC's optimization problem to prevent dangerous maneuvers.

What makes PECTS novel is that it learns both the system's stochastic dynamics and the appropriate safety barriers (CBFs) simultaneously using neural networks, specifically probabilistic and Lipschitz-bounded networks. This allows the AI to understand and account for uncertainty in its world model. During planning, PECTS uses a sampling-based optimizer that proactively discards any predicted trajectories that violate the learned safety constraints. The researchers validated PECTS in simulation studies, where it demonstrated superior performance over existing baseline methods, successfully maintaining safety under model uncertainty. This represents a significant step toward deploying RL agents in high-stakes physical applications like autonomous driving or robotic surgery, where a single unsafe action can have catastrophic consequences.