Learning Vision-Based Neural Network Controllers with Semi-Probabilistic Safety Guarantees

Xinhang Ma, Junlin Wu, Hussein Sibai, Yiannis Kantaros, Yevgeniy Vorobeychik · 2025

Read Learning Vision-Based Neural Network Controllers with Semi-Probabilistic Safety Guarantees by Xinhang Ma, Junlin Wu, Hussein Sibai, Yiannis Kantaros, Yevgeniy Vorobeychik (2025) 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-2503-00191.txt.

Description

Xinhang Ma and four co-authors teach camera guided controllers that come with mathematical safety assurances, in this 2025 paper. A controller that steers by images faces a hard verification problem: pixels are high dimensional, and how they relate to the true state is unknown. Most learned controllers offer no formal guarantee at all.

Their framework mixes reachability analysis with a conditional generative model of images and distribution free tail bounds, a combination they call semi probabilistic verification. Training then uses a safety shaped loss, a sampler that hoards the most instructive near failure examples, and a curriculum. Tests spanned flight simulator landings, simulated and real miniature car lane following, and drone obstacle avoidance. In each, the learned controllers kept their safety bounds while performing strongly on the task itself.

Catalog record

CREATORXinhang Ma, Junlin Wu, Hussein Sibai, Yiannis Kantaros, Yevgeniy Vorobeychik
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE30 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2503-00191

Browse the full library · About the Tower of Babel