Abstract:Safety-critical control tasks with high levels of uncertainty are becoming increasingly common. Typically, techniques that guarantee safety during learning and control utilize constraint-based safety certificates, which can be leveraged to compute safe control inputs. However, excessive model uncertainty can render robust safety certification methods or infeasible, meaning no control input satisfies the constraints imposed by the safety certificate. This paper considers a learning-based setting with a robust safety certificate based on a control barrier function (CBF) second-order cone program. If the control barrier function certificate is feasible, our approach leverages it to guarantee safety. Otherwise, our method explores the system dynamics to collect data and recover the feasibility of the control barrier function constraint. To this end, we employ a method inspired by well-established tools from Bayesian optimization. We show that if the sampling frequency is high enough, we recover the feasibility of the robust CBF certificate, guaranteeing safety. Our approach requires no prior model and corresponds, to the best of our knowledge, to the first algorithm that guarantees safety in settings with occasionally infeasible safety certificates without requiring a backup non-learning-based controller.
What problem does this paper attempt to address?
### Problems the paper attempts to solve
The paper aims to solve the safety assurance problem in control tasks with high uncertainty and safety requirements. Specifically, the paper focuses on how to restore the feasibility of the control barrier function (CBF) constraint through online learning methods to ensure system safety when the initial system dynamics model is unknown and safety certificates are not feasible.
### Background and problem setting
1. **System model**:
- Consider a control - affine system:
\[
\dot{x} = f(x)+g(x)u
\]
where \(x\in X\subseteq\mathbb{R}^n\) and \(u\in U\subseteq\mathbb{R}^m\), \(f:\mathbb{R}^n\rightarrow\mathbb{R}^n\) and \(g:\mathbb{R}^n\rightarrow\mathbb{R}^{n\times m}\) are partially unknown locally Lipschitz continuous functions, representing the drift dynamics and input matrix respectively.
2. **Safety definition**:
- The safe set \(C\) is defined as the super - level set of a known continuously differentiable function \(h:\mathbb{R}^n\rightarrow\mathbb{R}\):
\[
C\triangleq\{x\in\mathbb{R}^n\mid h(x)\geq0\}
\]
- The safety of the system means that \(C\) is forward - invariant under the closed - loop system, that is, if \(x(0)\in C\), then for all \(t\geq0\), \(x(t)\in C\).
3. **Control barrier function (CBF)**:
- CBF is a tool for ensuring system safety, which is defined as:
\[
\sup_{u\in U}\dot{h}(x, u)\triangleq\sup_{u\in U}\frac{\partial h}{\partial x}(x)(f(x)+g(x)u)\geq-\alpha(h(x))
\]
where \(\alpha\in\mathcal{K}_\infty\) is an extended class \(\mathcal{K}_\infty\) function.
### Main contributions
1. **Online learning algorithm**:
- An online learning algorithm based on the Gaussian Process (GP) is proposed, which can ensure system safety without a prior model or a backup safety controller.
- When the CBF constraint becomes infeasible, the algorithm uses Bayesian optimization techniques to quickly learn the time derivative of the CBF to restore the feasibility of the CBF constraint as soon as possible.
2. **Theoretical guarantee**:
- It is proved that if the sampling frequency is high enough, the algorithm can restore the feasibility of the CBF constraint, thus ensuring system safety.
- This is the first online learning algorithm that can ensure safety without a prior model or a backup safety control law.
### Numerical verification
1. **Cruise control system**:
- The cruise control system is used for numerical verification, showing the ability of the algorithm to ensure safety by efficiently learning the model online without a prior model.
2. **Quadrotor aircraft**:
- Numerical simulations of the quadrotor aircraft are carried out to further verify the effectiveness of the algorithm.
### Conclusion
The paper proposes a new online learning algorithm, which can restore the feasibility of the CBF constraint by efficiently learning the system model to ensure system safety when the initial system dynamics model is unknown and safety certificates are not feasible. This method is...