H2-EMV: Hierarchical Memory Framework for Lifelong Robot Deployment
Researchers have introduced a new framework called H2-EMV that enables humanoid robots to decide what information to keep based on how users interact with them. This system gradually builds a hierarchical episodic memory and uses a language model to identify what can be forgotten, adjusting its rules based on user feedback about what to omit. When tested in simulated home environments and with 20.5 hours of real data from ARMAR-7, H2-EMV succeeded in cutting memory size by 45% and reducing query-time computation by 35%, all while preserving accuracy. This development addresses the challenge of managing extensive episodic memories from continuous multimodal inputs, allowing robots to answer questions like "Where did you put my keys?" by adapting to user preferences.
Key facts
- H2-EMV is a framework for lifelong robot deployment
- It uses hierarchical episodic memory
- Selective forgetting is based on language-model relevance estimation
- Rules are updated via user feedback
- Tested on simulated household tasks and 20.5-hour ARMAR-7 recordings
- Reduces memory size by 45%
- Reduces query-time compute by 35%
- Maintains question-answering accuracy
Entities
Institutions
- arXiv