Abstract: While control barrier functions are employed in addressing safety, control synthesis methods based on them generally rely on accurate system dynamics. This is a critical limitation, since the dynamics of complex systems are often not fully known. Supervised machine learning techniques hold great promise for alleviating this weakness by inferring models from data. We propose a novel control barrier function-based framework for safe control through event-triggered learning, which switches between prioritizing control performance and improving model accuracy based on the uncertainty of the learned model. By updating a Gaussian process model with training points gathered online, the approach guarantees the feasibility of control barrier function conditions with high probability, such that safety can be ensured in a data-efficient manner. Furthermore, we establish the absence of Zeno behavior in the triggering scheme, and extend the algorithm to sampled-data realizations by accounting for inter-sampling effects. The effectiveness of the proposed approach and theory is demonstrated in simulations.
Introduction
In safety-critical control, Control Barrier Functions (CBFs) are a popular tool used to ensure a system stays within a "safe set." However, CBFs traditionally require a perfect mathematical model of the system's physics. In the real world, dynamics are often too complex to model analytically. While Gaussian Processes (GPs) can be used to learn unknown dynamics from data, combining them with CBFs creates two major challenges: optimization infeasibility when uncertainty is high, and "Zeno behavior," where online learning triggers an infinite number of updates in a finite amount of time.
Methodology
The proposed framework bridges this gap using a multi-layered approach:
- GP-based Safety Filter: The control problem is formulated as a Second-Order Cone Program (SOCP). Unlike standard quadratic programs, this formulation explicitly accounts for the probabilistic uncertainty bounds provided by the GP.
- Event-Triggered Learning: Instead of updating the model at fixed time intervals, the system only triggers a "learning event" when a specific condition—based on the ratio of model mean to uncertainty—is violated.
- Safe Excitation Signal: When a learning event is triggered, the controller applies a specific, finite control input designed to maximize information gain in a "safe direction." This ensures the model improves exactly where it needs to for the controller to remain mathematically feasible.
- Recursive Feasibility and Zeno Avoidance: The authors provide rigorous proofs that if the system starts safe, it will stay safe (with high probability) and that the time between updates will always be greater than zero, ruling out Zeno behavior.
Experimental Results
The method was evaluated using a high-fidelity numerical simulation of Adaptive Cruise Control (ACC). An "ego" vehicle must maintain a safe distance from a lead vehicle while trying to reach a desired speed, with rolling resistance treated as an unknown function.
As shown in the simulation results, the event-triggered approach maintained safety and feasibility for the entire 30-second run using only 50 data points. In contrast, a standard time-triggered approach became infeasible and violated safety constraints after just 6 seconds.
Furthermore, the proposed method successfully avoided Zeno behavior. While previous state-of-the-art methods suffered from an explosion in update frequency that would halt hardware simulations, the new method maintained a positive "inter-event time" throughout the mission.
Conclusion
This research presents a significant step forward in making safety-critical autonomous systems more robust in uncertain environments. By strategically injecting excitation signals only when model uncertainty threatens feasibility, the framework provides probabilistic safety guarantees in a data-efficient manner. Future work will explore scaling GP inference for long-duration missions and scenarios where the definition of safety itself might be uncertain.
Summary updated