Reproduce “CoRA-VTLA: Contact-Guided Attention and Queryable Response Memory for Tactile Post-Training” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Chengye Song, Rui Song, Shengqi Wang, Jiawei Zhang, Hongzhou Wang, Xueyang Guo, Yang Liu
CoRA-VTLA adds response-aware tactile memory and image-grounded contact tokens to a pretrained vision-language-action policy for manipulation after contact. Its memory retrieves recent pressure and wrench changes using current contact and robot state, while gaze supervision directs a semantic token toward fingertip regions. Across three MuJoCo tasks, it achieves 73.33% mean success versus 50% for direct fusion; this does not isolate either mechanism.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.