Hide-and-Seek in Trajectories: Discovering Failure Signals for VLA Runtime Monitoring
Seongheon Park, Wendi Li, Changdae Oh, Samuel Yeh, [0.2em] Zsolt Kira, Michael Hagenow, Sharon Li, [0.2em] University of Wisconsin–Madison · 2026
Read Hide-and-Seek in Trajectories: Discovering Failure Signals for VLA Runtime Monitoring by Seongheon Park, Wendi Li, Changdae Oh, Samuel Yeh, [0.2em] Zsolt Kira, Michael Hagenow, Sharon Li, [0.2em] University of Wisconsin–Madison (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-2605-30834.txt.
Description
Park and colleagues propose Hide-and-Seek, framing VLA failure detection as a learning problem. VLA models follow orders and stretch across tasks, but remain open to faults that harm trust. Old methods either rely on costly resampling or outside models, while others spread labels evenly, hiding local signals. By combining inter- and intra-trajectory objectives, Hide-and-Seek localizes failure actions and induces signals from trajectory-level supervision alone, without step-level annotation. Evaluated on LIBERO, VLABench, and a real robot platform, the method achieves top failure spotting.
Catalog record
| CREATOR | Seongheon Park, Wendi Li, Changdae Oh, Samuel Yeh, [0.2em] Zsolt Kira, Michael Hagenow, Sharon Li, [0.2em] University of Wisconsin–Madison |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 93386 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2605.30834