Event

EventSync proposes event-driven memory update framework

1 reports · 1 sources

Researchers including Jingjiao Zhao proposed a hierarchical framework, EventSync, which uses event-driven memory updates to coordinate task reasoning and action execution in robot manipulation, for history-dependent tasks where the next action cannot be determined from current observations alone. In the framework, a vision-language model generates subgoals based on the task plan and interaction history, a vision-language-action model handles execution, and a lightweight Event Monitor combines visual observations with end-effector trajectories to judge whether a subgoal is complete; only after completion evidence is obtained is the completed subgoal written into the history and subsequent reasoning triggered. The paper reports that, under the setting where the executor has visual history, EventSync achieves an overall success rate of 64.5% on the RoboMME benchmark, 20.0 percentage points higher than the best baseline reported on that benchmark; the researchers state that experiments on four real-robot tasks with differing memory requirements verified the framework's effectiveness in real manipulation.