Wonje Jeung, Sangyeon Yoon, Hyesoo Hong, Yoonjun Cho, Dongjae Jeon, Bumjun Kim, Jean Oh, Youngjae Yu and Albert No · The full paper is marked CC BY 4.0, and the Figure 1 caption contains no separate third-party credit or license exception. CC BY 4.0 permits commercial reproduction and sharing with creator attribution, a link to the licensed material and license, an indication of modifications, and no implication of endorsement. The figure should be reproduced without misleading alteration.
ANALYSIS Robotics
RoboRMBench Exposes Wording-Sensitive Reward Scores in Robot Learning
A new preprint tests vision-language reward models on 2,390 recorded real-robot trajectories and finds that equivalent instructions can produce contradictory judgments, revealing a measurable reliability gap before physical deployment.