Curvature-Guided Safety Filters: State-Dependent Hessian-Weighted Projection with Provable Performance Bounds
Ziyan Lin, Liang Xu · 2026
Read Curvature-Guided Safety Filters: State-Dependent Hessian-Weighted Projection with Provable Performance Bounds by Ziyan Lin, Liang Xu (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2602-12603.txt.
Description
Safety filters keep learning-enabled controllers within safe states and inputs, but common Euclidean projections onto the safe set ignore long-term performance. Ziyan Lin led this work with one co-author, Liang Xu, giving a state-dependent, Hessian-guided projection for safety filtering. It preserves convexity while improving performance, unlike direct value optimisation, which can be nonconvex.
The key idea picks a weighted projection matrix from curvature of the action-value function. It biases corrections toward action directions with higher value sensitivity. Two results are established: a uniform bound on the performance gap to the safe value-optimal action, and a condition under which weighted projection beats Euclidean projection in long-term value. A data-driven construction via Q-function learning supports black-box controllers. Quadrotor tracking-and-avoidance simulations show safety maintained with less value loss than Euclidean projection, at real-time-compatible cost.
Catalog record
| CREATOR | Ziyan Lin, Liang Xu |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 40 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2602-12603