Event
Proposes RePlanMem-VLA visual memory method
1 reports · 1 sources
Victoria Smelova et al. propose RePlanMem-VLA on arXiv, turning the VLA policy's own plan revisions into an online memory-write signal: during receding-horizon execution, if predictions of the same future action by adjacent action chunks diverge abnormally, this is treated as a task-relevant event, triggering a bounded visual memory update, and a robot-mask write filter is used to suppress redundant frames. The paper reports that the event detector achieves a ROC-AUC of 0.981–0.997; in the evaluated memory-retention experiments, RePlanMem-VLA retained task-relevant observations until subsequent decision points in 100% of episodes; real-robot replay data also show that the effects of event selection and redundancy suppression extend beyond simulation.
Report timeline
- RePlanMem-VLA: Guiding VLA Keyframe Visual Memory with Plan Revision Signals
An arXiv paper proposes RePlanMem-VLA (Replanning-Guided Keyframe Memory for VLA), which turns a policy's own plan revisions into an online memory-write signal: during receding-horizon execution, if consecutive action chunks show unusually large disagreement in their predictions of the same future action, this is treated as a task-relevant event.
arXiv cs.RO ·