Proactive robots are designed to step in before a user asks, but working out what someone is about to do is tricky. The abstract of a new arXiv paper notes that a straightforward route — inferring intentions from commonsense knowledge — can fail because the same observed behavior can imply multiple possible intentions. Without additional context, the robot may guess wrong.

The paper outlines an agentic framework that gives robots a procedural assistance memory. Rather than relying solely on general commonsense, the robot draws on stored procedures or past assistance patterns to interpret current behavior and decide what help is appropriate. This memory-based layer is intended to make anticipation more robust in real-world situations where actions are ambiguous.

Because the source is only an abstract, details of the architecture and evaluation are not available here. The significance lies in the shift from static intention inference to a memory-augmented, agentic approach — a step toward robots that learn how to be helpful from experience.