Towards safety-critical control of autonomous systems

This thesis addresses these challenges by developing novel learning-based safety-critical control frameworks aimed at enhancing both the autonomy and trustworthiness of complex autonomous systems operating under uncertainty. Safety in this context specifically refers to the formal guarantee that trajectory tracking controllers consistently maintain system states within predefined safe operational limits, mathematically defined through barrier functions and constraint sets, avoiding hazardous states such as collisions, control saturation, or violations of physical limitations and critical performance constraints.

This is achieved through control barrier function formulations that provide provable forward invariance of safe sets. Trustworthiness in this context is defined as the verifiable assurance that these systems consistently maintain safe, reliable, and predictable trajectory tracking behaviour, encompassing both mathematical guarantees of safety constraint satisfaction and demonstrated reliability through experimental validation, thereby mitigating key obstacles to their broader acceptance and deployment.

The primary novel contributions of this thesis lie in developing adaptive, safety-critical trajectory tracking controllers. The initial part introduces innovative control techniques, including novel applications of Barrier Lyapunov Functions and adaptive Control Barrier Functions, designed to rigorously enforce safety constraints while ensuring accurate trajectory tracking and optimising overall system performance. Unlike existing approaches, this work uniquely combines these safety mechanisms with learning-based parameter adaptation, enabling real-time responsiveness to uncertainty.

Further contributions involve a distributed Model Predictive Control strategy that enables cooperative coordination among multiple autonomous agents. This approach is experimentally validated on underwater robotic platforms operating in realistic, dynamic settings, demonstrating practical viability. Finally, reinforcement learning methodologies are employed for the real-time adaptation of control parameters, yielding enhanced autonomy and flexibility without compromising critical safety guarantees.

Collectively, these contributions represent a meaningful step forward in enhancing the autonomy, adaptability, and trustworthiness of autonomous systems. By integrating rigorous control theory with advanced machine learning techniques, this work contributes to the ongoing development of reliable and safe autonomous technologies capable of operating in complex, real-world environments.