New Publication in Computers & Chemical Engineering: Safe Reinforcement Learning through Adaptive Model Predictive Shielding

The work of Hilde Gerold under the supervision of Sergio Lucia, titled “Safe reinforcement learning via adaptive robust model predictive shielding”, has been published in the March 2026 issue of Computers & Chemical Engineering (Volume 206).

Reinforcement learning (RL) has shown significant potential for advanced control applications, particularly in complex and nonlinear systems. However, ensuring safe operation during deployment remains a major challenge, especially when system constraints must be strictly satisfied despite uncertainty and limited model knowledge.

In this work, the authors address this challenge through the concept of model predictive shielding. In this framework, proposed control actions generated by a learning-based controller are continuously verified using a predictive system model and forward simulations. Whenever a future constraint violation is detected, the unsafe action is replaced by a safe backup action before it can be applied to the system.

The proposed approach extends existing shielding methods in two important directions. First, the authors introduce an adaptive safety parameter that dynamically adjusts the level of conservativeness depending on the uncertainty in the system. This reduces unnecessary interventions by the backup controller while still maintaining safety guarantees. Second, the work incorporates an approximate robust nonlinear model predictive controller as the backup policy, enabling fast and reliable safe actions that are suitable for real-time implementation.

By combining adaptive robustness with efficient predictive control, the proposed framework improves both the safety and performance of reinforcement learning-based controllers and represents an important step toward the deployment of learning-based control methods in safety-critical industrial applications.

You can read the full article here.