Reproduce “ATOM-CBF: Adaptive Safe Perception-Based Control under Out-of-Distribution Measurements” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Kai S. Yun, Navid Azizan
The ATOM-CBF framework enables adaptive safe control for autonomous systems utilizing learned perception modules by dynamically adjusting safety guarantees based on real-time epistemic uncertainty from out-of-distribution measurements. This approach achieved a 0% collision rate in high-fidelity simulations across F1Tenth vehicles and quadruped robots operating in challenging, unseen environments.
Key insights
- Dynamically adapting safety guarantees based on real-time epistemic uncertainty allows systems to maintain safety in OoD conditions without requiring online ground-truth data or prior distribution shift information.
- A critical trade-off exists between the strength of epistemic uncertainty quantification (EUQ) and the resulting conservatism of the safety controller, where highly sensitive EUQ can lead to over-conservative behavior (deadlocks).
- A calibrated filtering procedure for the base error ratio ("phi_cal") is essential to balance safety assurances with task performance, preventing excessive conservatism.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.