ORN-CBF: Learning Observation-Conditioned Residual Neural Control Barrier Functions Via Hypernetworks
Bojan Derajic, Sebastian Bernhard, Wolfgang Hoenig
AI summary
Problem
Designing safe control filters for autonomous systems in unknown, partially observable environments remains difficult because offline CBF methods cannot adapt in real-time, while existing learning-based approaches lack rigorous safety guarantees and fail to recover maximal safe sets.
Approach
The method uses a hypernetwork to condition a lightweight main network on real-time environment observations, learning only the residual component of a Hamilton-Jacobi value function to guarantee the safe set never overlaps with observed obstacles.
Key results
- Novel observation-conditioned neural CBF guaranteeing non-intersection with failure sets
- Efficient hypernetwork architecture enabling real-time safety filtering
- Improved success rates and out-of-domain generalization in simulation and hardware
- Local patch training strategy reducing memory and computational overhead
Why it matters
Enables reliable, safe navigation for autonomous robots in dynamic environments by combining rigorous safety guarantees with real-time adaptability, benefiting robotics and autonomous systems researchers and practitioners.
Abstract
Control barrier functions (CBFs) have been demonstrated as an effective method for safety-critical control of autonomous systems. Although CBFs are simple to deploy, their design remains challenging, motivating the development of learning-based approaches. Yet, issues such as suboptimal safe sets, applicability in partially observable environments, and lack of rigorous safety guarantees persist. In this work, we propose observation-conditioned neural CBFs based on Hamilton-Jacobi (HJ) reachability analysis, which approximately recover the maximal safe sets. We exploit certain mathematical properties of the HJ value function, ensuring that the predicted safe set never intersects with the observed failure set. Moreover, we leverage a hypernetwork-based architecture that is particularly suitable for the design of observation-conditioned safety filters. The proposed method is examined both in simulation and hardware experiments for a ground robot and a quadcopter. The results show improved success rates and generalization to out-of-domain environments compared to the baselines.