Reproduce “EvoMemNav: Efficient Self-Evolving Fine-Grained Memory for Zero-Shot Embodied Navigation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Zuhao Ge, Xiaosong Jia, Chao Wu, Yuchen Zhou, Zuxuan Wu, Yu-Gang Jiang
EvoMemNav introduces a self-evolving, fine-grained memory framework for zero-shot embodied navigation that effectively addresses memory limitations by combining detailed visual evidence with structured semantic organization and a training-free adaptation mechanism. It achieved a 59.6% Success Rate and 38.9% SPL on GOAT-Bench Val Unseen, while reducing VLM calls by 41% compared to a baseline.
Key insights
- Integrating fine-grained visual evidence directly into a structured, hierarchical memory graph (VSMGraph) allows for more robust disambiguation and VLM reasoning than symbolic or dense 3D maps.
- A two-stage navigation policy that budgets VLM queries by first compressing candidate sets and then performing targeted verification on a smaller set significantly improves efficiency and reduces errors like premature stopping.
- Allowing the memory system to continually adapt and learn from past experiences without explicit retraining, through reflection-driven updates of goal-conditioned priors, enhances lifelong learning capabilities and cumulative performance.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.