William D. Compton,Max H. Cohen,Aaron D. Ames
Abstract:Safety filters leveraging control barrier functions (CBFs) are highly effective for enforcing safe behavior on complex systems. It is often easier to synthesize CBFs for a Reduced order Model (RoM), and track the resulting safe behavior on the Full order Model (FoM) -- yet gaps between the RoM and FoM can result in safety violations. This paper introduces \emph{predictive CBFs} to address this gap by leveraging rollouts of the FoM to define a predictive robustness term added to the RoM CBF condition. Theoretically, we prove that this guarantees safety in a layered control implementation. Practically, we learn the predictive robustness term through massive parallel simulation with domain randomization. We demonstrate in simulation that this yields safe FoM behavior with minimal conservatism, and experimentally realize predictive CBFs on a 3D hopping robot.
What problem does this paper attempt to address?
The problem that this paper attempts to solve is: the challenges encountered when using control barrier functions (CBFs) to ensure safe behavior in complex nonlinear systems. Specifically, CBFs are usually synthesized on a reduced - order model (RoM) and then these safe behaviors are tracked on a full - order model (FoM). However, the differences between RoM and FoM may lead to safety violations. To solve this problem, this paper introduces predictive control barrier functions (Predictive CBFs), defines a robustness term through FoM simulation prediction, and adds it to the CBF conditions of RoM, thereby ensuring safety in hierarchical control implementation.
### Specific description of the problem
1. **Background and challenges**:
- Control barrier functions (CBFs) are effective tools for ensuring the safe behavior of complex systems.
- In high - dimensional or complex nonlinear systems, it is very difficult to directly synthesize CBFs for the full - order model (FoM).
- The common practice is to first synthesize CBFs on a reduced - order model (RoM) and then track these safe behaviors on the full - order model.
- However, the differences between RoM and FoM may lead to safety violations, especially in high - dimensional or complex systems.
2. **Limitations of existing methods**:
- Existing methods rely on strict tracking assumptions (such as exponential tracking), which are difficult to meet in practical applications, resulting in conservative behavior.
- It is also very difficult to calculate the certificates of these tracking assumptions, and the results are often overly conservative approximations.
3. **The method proposed in the paper**:
- Introduce predictive control barrier functions (Predictive CBFs), define a robustness term (robustness term) through FoM simulation prediction, and add it to the CBF conditions of RoM.
- This robustness term is learned through large - scale parallel simulation and domain randomization to ensure the safety of FoM with minimal conservatism.
- It is theoretically proven that this method can ensure safety, and its application effect on a 3D jumping robot is demonstrated in the experiment.
### Formula representation
- **Dynamic equation of RoM**:
\[
\dot{z} = f(z)+g(z)v
\]
where \(z\in Z\subseteq\mathbb{R}^n\), \(v\in V\subset\mathbb{R}^m\).
- **Dynamic equation of FoM**:
\[
\dot{x} = F(x)+G(x)u
\]
where \(x\in X\subseteq\mathbb{R}^N\), \(u\in U\subseteq\mathbb{R}^M\).
- **Safety condition of predictive CBF**:
\[
L_fh(z)+L_gh(z)v\geq -\alpha(h(z))+\delta(x)
\]
where \(\delta(x)\) is the robustness term obtained through prediction.
### Summary
This paper solves the safety problems that may be encountered when synthesizing CBFs using a reduced - order model in complex nonlinear systems by introducing predictive control barrier functions (Predictive CBFs). By learning the robustness term through large - scale parallel simulation and domain randomization, this method can ensure safety while reducing conservatism and is applicable to a wider range of complex systems.