arXiv Robotics — research abstracts· Koya Sakamoto, Daichi Azuma, Shuhei Kurita, Naoya Chiba, Yusuke Iwasawa, Yutaka Matsuo, Taiki Miyanishi·· 10 days agoEditorial score58
Rendering-Free Lookahead for Question-Guided Active Vision
Rendering-Free Lookahead for Question-Guided Active Vision
Summary
This research proposes Rendering-Free Lookahead (RFL), a policy for selecting camera motions in robotics to answer viewpoint-dependent questions. RFL uses a vision-language model to estimate the usefulness of future views during training and selects motions at deployment without rendering future scenes, achieving a 43% improvement in performance on unseen environments.
Source: arXiv Robotics — research abstracts · Read original article ↗
Loading article text…
Source：arXiv Robotics — research abstracts · arxiv.org