Research map / Multi-agent systems and games / Human-AI and goal-conditioned
MemoNav: Working Memory Model for Visual Navigation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes MemoNav, a working-memory-inspired topological-map model for image-goal navigation that forgets low-attention nodes, adds a global LTM node, and encodes STM+LTM with GATv2 into working memory.
Problem
Image-goal navigation in unfamiliar environments; existing topological-map methods use all historical node features for decision-making, causing inefficient exploration and poor multi-goal performance.
Approach
Combines three memory types: short-term memory (topological map node features) with attention-based selective forgetting, a zero-initialized global node aggregating scene-level long-term memory, and a GATv2 graph attention module that fuses retained STM and LTM into a working memory used by Transformer decoders and a policy.
Key results
- On Gibson 1-goal, MemoNav reaches 74.7% SR and 57.9% SPL versus 70.0% SR and 55.4% SPL for VGM.
- On Gibson multi-goal, MemoNav improves 4-goal progress to 20.1 PR vs 17.1 for VGM (PR 38.0 vs 29.5 at 2-goal... 3-goal PR 50.8 vs 42.9).
- Ablation shows forgetting, LTM and WM each help, and the full combination gives the best SR/PR across all difficulty levels.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use MemoNav instead of VGM | ImageNav · Gibson 4-goal | PR: 21.5 -> 28.9 | Helped |
| Use MemoNav instead of VGM | ImageNav · Matterport3D 3-goal | PR: 11.8 -> 13.6 | Helped |
| Add the forgetting module that excludes low-attention nodes (ablation) | MemoNav · Gibson 1-goal | SR: 52.1 -> 55.1 | Helped |
| Add the forgetting module that excludes low-attention nodes (ablation) | MemoNav · Gibson 2-goal | PR: 42.9 -> 44.9 | Helped |
| Add LTM to continuously fuse STM features (ablation) | MemoNav · Gibson 1-goal | SR: 52.1 -> 58.9 | Helped |
| Add both the forgetting module and LTM (ablation) | MemoNav · Gibson 2-goal | PR: 42.9 -> 48.1 | Helped |
| Add graph-attention working memory to the forgetting-plus-LTM model (ablation) | MemoNav · Gibson 3-goal | PR: 37.5 -> 38.0 | Helped |
| Include LTM in working memory (ablation) | MemoNav · Gibson 2-goal | PR: 46.3 -> 50.8 | Helped |
| Use a learned LTM rather than a randomly selected STM feature (ablation) | MemoNav · Gibson 2-goal | PR: 45.9 -> 50.8 | Helped |
| Use a learned LTM rather than averaged STM features (ablation) | MemoNav · Gibson 2-goal | PR: 46.1 -> 50.8 | Helped |
Surprising findings
- Forgetting 80% of short-term memory still keeps high success on 3-goal tasks, with only slight drops.
- Replacing LTM with an averaged STM feature decreases performance less on harder 3- and 4-goal tasks than on easier ones.
Limitations
- Failure cases include stopping mistakenly, missing the goal, not close enough, and over-exploring.
- Does not use object semantics, which the authors note as orthogonal to their contributions.
- Forgetting threshold p is set empirically to 20% and trained without forgetting.
Experimental setup
Datasets: Gibson Matterport3D Gibson 1-goal episodes (public dataset of 14 unseen scenes)
Benchmarks: Gibson 1-goal Gibson multi-goal (2-4 goals) Matterport3D multi-goal (1-4 goals)
Metrics: SR SPL PR PPL
Baselines: CNNLSTM ANS NTS TSGM VGM
Builds on: Visual Graph Memory (VGM) GATv2 PPO ETC Longformer
Closest papers by meaning
- EfficientNav: Towards On-Device Object-Goal Navigation with Navigation Map Caching and Retrieval (NeurIPS 2025 · similarity 0.89)
- 3D-Mem: 3D Scene Memory for Embodied Exploration and Reasoning (CVPR 2025 · similarity 0.89)
- Change-Robust Online Topological Memory for Long-Term Relocalization and Semantic Navigation (NeurIPS 2026 · similarity 0.88)
- C-NAV: Towards Self-Evolving Continual Object Navigation in Open World (NeurIPS 2025 · similarity 0.87)
- LangMap: A Human-Verified Benchmark for Hierarchical Open-Vocabulary Goal Navigation (NeurIPS 2026 · similarity 0.87)
- Imagine Before Go: Self-Supervised Generative Map for Object Goal Navigation (CVPR 2024 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.