Learning with Object-centric Representations of Tactile Interactive Perception for Robot Manipulation
Learns object representations from tactile exploration, aligned with text descriptions of properties, and uses them as context for manipulation. Reports task success from 41% to 92% (seen) and 19% to 67% (unseen).通过触觉探索学习物体表征，并与物理属性的文本描述对比对齐，再作为语境指导操作策略。报告任务成功率在已见物体上从 41% 升到 92%，未见物体从 19% 升到 67%。
BibTeX
@misc{yang2026learning,
title = {{Learning with Object-centric Representations of Tactile Interactive Perception for Robot Manipulation}},
author = {Xinyi Yang and Zilin Si and Zhuowei Xu and Zeynep Temel and Oliver Kroemer},
year = {2026},
eprint = {2609.33235},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2609.33235},
}