Adaptive and Learning-Based Control of Safety-Critical Systems

M. Cohen, Calin BeltaSynthesis 2023 · DOI · OpenAlex

Abstract

As autonomous systems transition from controlled environments to the real world, they must navigate significant physical uncertainties while maintaining strict safety requirements. This work presents a unified framework that synthesizes Control Barrier Functions (CBFs) with adaptive control and machine learning. By treating safety as a problem of set invariance, the authors demonstrate how systems can perform online model adaptation and optimization without compromising formal safety guarantees.

Introduction

The deployment of autonomous agents—such as drones, self-driving vehicles, and robotic assistants—in dynamic environments necessitates a control paradigm that is both adaptive and safe. While Reinforcement Learning (RL) excels at optimizing performance, it frequently lacks the formal guarantees required for safety-critical applications. Conversely, traditional robust control methods often fail to account for evolving system dynamics. This paper addresses the gap by providing a framework for online adaptation with formal correctness, ensuring that systems remain within safe operating boundaries even as they learn from new data.

Methodology

The proposed framework integrates several advanced mathematical and control concepts to achieve safe autonomy:

Control Barrier Functions (CBFs)

CBFs serve as the primary mechanism for safety. Unlike Lyapunov functions, which focus on reaching a goal state (stability), CBFs are designed to ensure that the system state never exits a predefined "safe set."

Adaptive Control and Concurrent Learning

The authors extend traditional adaptive control to safety-critical contexts through Adaptive Control Barrier Functions (aCBFs). To enhance the speed and accuracy of model updates, Concurrent Learning is employed. This technique leverages both current and historical data, allowing the system to identify physical parameters (e.g., mass, friction) without the need for the "Persistence of Excitation" typically required in classical adaptive methods.

Safe Model-Based Reinforcement Learning (MBRL)

In the RL domain, the framework introduces a "safety filter" based on CBFs. This filter monitors the actions proposed by an RL agent and overrides them if they threaten to violate safety constraints, allowing for safe exploration and optimization.

Formal Logic Integration

For complex, multi-stage missions, the framework utilizes Linear Temporal Logic (LTL). High-level logic requirements are decomposed into a sequence of "reach-avoid" problems, which are then solved using the low-level safety controllers.

Results and Insights

The framework's efficacy is demonstrated through rigorous mathematical proofs and numerical simulations.

  • Safety as Set Invariance: The core contribution is the mathematical proof that if a valid aCBF exists, the system is guaranteed to remain within the safe region despite initial model errors.
  • Dynamic Safety Bounds: The system manages the trade-off between learning and safety by using "worst-case bounds." As more data is collected and model confidence increases, these bounds shrink, allowing the controller to become more efficient and less conservative.
  • Real-Time Feasibility: The control laws are formulated as Quadratic Programs (QPs). These optimization problems can be solved in milliseconds, making the framework suitable for high-frequency deployment on physical hardware.

Conclusion and Future Work

This framework provides a robust foundation for safe, adaptive autonomy. However, several challenges remain. The current methods largely depend on knowing the underlying structure of the system's physics and assume full state observability. Future research will focus on handling "black-box" uncertainties where the model structure is unknown and integrating robust state estimation to account for sensor noise in partially observable environments.

Summary based on "Adaptive and Learning-Based Control of Safety-Critical Systems" by Max Cohen and Calin Belta (2023).

Summary updated

BibTeX
@inproceedings{cohen2023adaptive,
  author = {M. Cohen and Calin Belta},
  title = {{Adaptive and Learning-Based Control of Safety-Critical Systems}},
  booktitle = {Synthesis},
  year = {2023},
  doi = {10.1007/978-3-031-29310-8},
}
Figure