Introduction
Ensuring safety in autonomous systems is paramount, yet real-world applications often involve noisy sensors and imperfect physical models. This paper introduces a framework to learn Robust Output Control Barrier Functions (ROCBFs) from safe expert demonstrations. By leveraging expert data and explicitly accounting for uncertainty, the authors provide a practical way to synthesize safety-guaranteed control laws that operate directly from high-dimensional inputs like camera images.
Problem Statement: Safety Under Uncertainty
Traditional safety methods, specifically Control Barrier Functions (CBFs), are powerful tools for ensuring a system stays within a "safe set" of states. However, most CBF approaches assume perfect state knowledge and a perfect model. In practice, systems rely on "partial observations"—estimating state from noisy sensors—and simplified physical models. When these assumptions fail, a standard "safe" controller might inadvertently lead to a collision.
The core challenge is learning a safety certificate that is robust to both perception errors and model uncertainties using only safe demonstration data.
The ROCBF Framework
The proposed approach bridges the gap between theory and practice through a two-step process:
- ROCBF Theory: The authors define the Robust Output Control Barrier Function, which modifies standard CBF conditions with "robustness terms." These terms act as a safety buffer, accounting for the maximum possible error in state estimation and system dynamics.
- Learning from Experts: Instead of manual engineering, the functions are learned via a constrained optimization problem. The algorithm finds a function $h(x)$ that remains positive for safe data and negative for "unsafe" data.
Key Algorithmic Innovations
- Boundary Point Detection: To avoid collecting actual crash data, the authors use a reverse k-nearest neighbors approach to identify the "edges" of safe expert trajectories. These boundary points are labeled as "unsafe" to provide contrast for learning.
- Convex Optimization: By representing the ROCBF as a linear combination of basis functions, the learning problem becomes convex, ensuring reliable and fast solutions.
Experimental Validation
The framework was validated using the CARLA autonomous driving simulator. Expert demonstrations were collected from a PID controller with full state access. For perception-based tasks, a ResNet-18 architecture estimated cross-track error from RGB dashboard camera images, providing a realistic, noisy output.
Key Results
- Robustness to Perception Noise: Using only RGB images, the ROCBF-based controller maintained safety despite significant estimation errors from the neural network "eyes."
- Performance Comparison: The robust approach significantly outperformed non-robust versions. While standard methods occasionally resulted in collisions under noisy perception, the ROCBF version maintained a wider safety margin and avoided crashes entirely.
Conclusion and Future Work
This work demonstrates that safety certificates can be effectively learned from behavior, even when the system lacks perfect state information. However, limitations remain: the theoretical guarantees rely on data density, and the method assumes known (or estimable) error bounds for sensors. Future research may focus on automatically estimating these uncertainty bounds online or extending the framework to complex multi-agent environments.
Summary updated