POLAR: Memory-Augmented Framework for Personalized Embodied AI Agents
A novel system named POLAR (Personalized Object Learning and Adaptive Retrieval) has been introduced to improve long-term personalization in multimodal large language model (MLLM)-powered embodied agents. Outlined in arXiv preprint 2605.26256, this framework tackles the issue of agents interpreting implicit user instructions based on previous interactions instead of standard commands. POLAR constructs a multimodal knowledge graph that encompasses both semantic memory (personalized context and visual concepts) and episodic memory (agent trajectories). It retrieves pertinent memories during task execution to understand current requests and direct actions. Evaluations across various machine learning benchmarks revealed enhanced performance in personalized assistance, marking a significant advancement in embodied AI by allowing agents to utilize user-specific context over time.
Key facts
- POLAR is a multimodal memory-augmented framework for personalized embodied agents.
- It organizes prior interactions into a multimodal knowledge graph.
- The knowledge graph includes semantic memory for personalized context and visual concepts.
- It also includes episodic memory for embodied experiences like agent trajectories.
- POLAR retrieves relevant memories to interpret current requests and guide task execution.
- The framework was evaluated across multiple machine learning benchmarks.
- It addresses the need for agents to leverage personalized context from long-term interactions.
- The work is published on arXiv with identifier 2605.26256.
Entities
Institutions
- arXiv