H²-EMV lets a humanoid learn what to remember. It builds a hierarchical episodic memory incrementally, forgets selectively using language-model relevance estimates conditioned on learned natural-language rules, and updates those rules from user feedback about forgotten details. On simulated household tasks and 20.5 hours of real recordings from ARMAR-7 it keeps question-answering accuracy while cutting memory size by 45% and query compute by 35%, and second-round accuracy rises by 70% as it adapts to a user’s priorities.
Relevance
Weeks of operation need forgetting, and here the forgetting policy is learned from the user. For long-term autonomy; the open question for field robots is who gives the feedback when no user is around.
Links
- arXiv
- Subfield: Semantic mapping, Robot learning