TaskAnchor improves robot manipulation by tracking task progress
TaskAnchor: Grounding Task State in Reactive VLAs for Long-Horizon Manipulation
Robotics
Summary
Robots often struggle to complete long tasks when similar-looking scenes require different actions depending on what step they are on. The authors introduce TaskAnchor, a way to help robots remember their progress by linking visual and language signals to the task history. This method lets robots better know what stage they are in without changing how they decide actions. TaskAnchor greatly improves success rates in robot tasks with only a tiny delay added per decision.
What this means in practice
- •For robotics engineers: Enable robots to better complete complex multi-step tasks by grounding decisions in task history without redesigning action systems.
- •For automation system developers: Improve industrial robotic systems' accuracy and reliability in long sequences by integrating task-state tracking with existing vision-language models.
Authors
Hengyan Liu, Wenlve Zhou, Bo Yue, Yongyi Su, Ruixiang Wang, Zhanqi Zhang, Dekun Lu, Wei Gao, Xiaofen Xing, Kui Jia
Abstract
Reactive vision--language--action (VLA) models struggle with long-horizon manipulation when visually similar observations can correspond to different actions depending on the task stage or interaction history. We refer to this ambiguity as task-state aliasing and introduce TaskAnchor, a lightweight adapter that grounds pretrained VLAs in execution history. TaskAnchor combines history-conditioned visual refinement with a milestone-supervised task-state coordinate, a scalar representing the semantic stage of execution. These signals are injected through the native visual and language interfaces, respectively, without introducing an explicit planner or modifying the action-generation mechanism. On RMBench, TaskAnchor achieves approximately 4.9--5.5$\times$ the average success rates of the published $π_{0.5}$ and X-VLA baselines, with consistent gains on RoboMemArena and real robots. The added latency is only 2.08\,ms per action chunk for $π_{0.5}$.