arXiv Robotics — research abstracts· Hengyan Liu, Wenlve Zhou, Bo Yue, Yongyi Su, Ruixiang Wang, Zhanqi Zhang, Dekun Lu, Wei Gao, Xiaofen Xing, Kui Jia·· 15 hours agoEditorial score62
TaskAnchor: Grounding Task State in Reactive VLAs for Long-Horizon Manipulation
TaskAnchor: Grounding Task State in Reactive VLAs for Long-Horizon Manipulation
Summary
This research introduces TaskAnchor, a lightweight adapter that improves reactive vision-language-action (VLA) policies for long-horizon manipulation by grounding task state through execution context injection. TaskAnchor achieves substantial performance gains over existing baselines in benchmarks and real-robot experiments, demonstrating reliable multi-stage execution with in-context cues from earlier human interactions.
Source: arXiv Robotics — research abstracts · Read original article ↗
Loading article text…
Source：arXiv Robotics — research abstracts · arxiv.org