This paper examines how memory-augmented vision-language-action (VLA) policies use time as a proxy for progress, leading to errors in counting tasks.
The research question addresses whether memory-augmented VLA policies accurately represent progress by remembering events or if they rely on elapsed time, which can misinterpret waiting as work.
The authors propose an experiment where robots perform counting tasks while briefly pausing or making null movements, observing how these actions affect the policy's memory and decision-making.
Evaluation on RoboMME tasks shows that memory indexed by time leads to 'phantom progress,' causing the robot to stop prematurely. The authors suggest that event-based indexing could mitigate this issue.
About this brief
This article summarizes a public preprint abstract. Its findings have not been independently reproduced for this publication. Consult the paper for methods, evaluation scope and limitations. Automated briefs may contain errors; the original source takes precedence.