Episodic Memory Banks for Lifelong Robot Learning: A Case Study Focusing on Household Navigation and Manipulation

Zichao Li

Keywords: episodic memory, lifelong learning, embodied AI, memory banks, robotic manipulation, sim-to-real transfer, hierarchical retrieval

TL;DR: Episodic Memory Banks enable lifelong learning for robots through hierarchical memory retrieval, achieving 84.4% accuracy across six benchmarks while reducing compute needs 3.6× and maintaining 72.4% real-world task success.

Abstract: This paper introduces Episodic Memory Banks, a novel architecture for lifelong learning in embodied agents that combines hierarchical memory retrieval with visual-language-action grounding. Our method achieves state-of-the-art performance across six benchmarks (HM3D, ALFRED, BEHAVIOR, Ego4D, DROID, RLBench), with 84.4% average memory retrieval accuracy (17.4% improvement over prior work). Key innovations include 3D-aware memory compression, contrastive sim-to-real alignment, and task-aware replay, enabling 72.4% real-world task success on DROID while using 3.6× fewer GPU resources than comparable approaches. Quantitative analysis reveals our method reduces catastrophic forgetting by 15.9% over six months and maintains sublinear computational scaling to 1M+ memory entries. The architecture's modular design permits integration with existing foundation models while providing interpretable memory access patterns.

Publishedpaper: No, never published elsewhere.

Submission Number: 17