Research map / Multi-agent systems and games / Human-AI and goal-conditioned
AstraNav-Memory: Contexts Compression for Long Memory
Junjun Hu, Xinda Xue, Botao Ren, Minghua Luo, Jintao Chen, Haochen Bai, Liangliang You, Mu Xu
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| compress visual tokens to about 30 per image with DINOv3 + PixelUnshuffle and keep long image history | Qwen2.5-VL-3B navigation policy · GOAT-Bench Val-Unseen | SR: 47.2 -> 62.7 vs MTU3D | Improved |
| same | Qwen2.5-VL-3B · HM3D-OVON Val-Unseen | SR: 40.8 -> 62.5 vs MTU3D* | Improved |
| apply 16x token compression to 50 stored images (ablation) | navigation policy · GOAT subset | train time per iter: 26.4 s -> 6.5 s | Improved |
| apply 16x token compression to 50 stored images (ablation) | navigation policy · GOAT subset | GPU memory: 90.6 GB -> 46.8 GB | Improved |
| apply 16x token compression to 50 stored images (ablation) | navigation policy · GOAT subset | SR: 60.2 -> 56.6 | Hurt |
| increase stored images from 50 to 100 at 16x compression (ablation) | navigation policy · GOAT subset | SR: 56.6 -> 57.5 | Improved |
| increase stored images from 100 to 200 at 16x compression (ablation) | navigation policy · GOAT subset | SR: 57.5 -> 55.2 | Hurt |
| use 4x compression with 100 images instead of 1x with 50 (ablation) | navigation policy · GOAT subset | SR: 60.2 -> 63.2 | Improved |
| use 64x compression instead of 16x (ablation) | navigation policy · GOAT subset | SR: 56.6 -> 42.5 (50 images) | Hurt |
| use DINOv3 ViT-M instead of ViT-S (ablation) | tokenizer backbone · GOAT subset | SR: 53.3 -> 57.5 | Improved |
| use DINOv3 ViT-L instead of ViT-M (ablation) | tokenizer backbone · GOAT subset | SR: 57.5 -> 59.4 | Improved |
| use DINOv3-based compressed encoder instead of original Qwen ViT (ablation) | navigation policy · per-category SR (Table 5) | SR: e.g. freezer 72 -> 88; carpet 61 -> 56 | Mixed |
Closest papers by meaning
- MemoNav: Working Memory Model for Visual Navigation (CVPR 2024 · similarity 0.90)
- C-NAV: Towards Self-Evolving Continual Object Navigation in Open World (NeurIPS 2025 · similarity 0.87)
- Memo: Training Memory-Efficient Embodied Agents with Reinforcement Learning (NeurIPS 2025 · similarity 0.87)
- Image2Sim: Scaling Embodied Navigation via Generative Neural Simulator (NeurIPS 2026 · similarity 0.86)
- 3D-Mem: 3D Scene Memory for Embodied Exploration and Reasoning (CVPR 2025 · similarity 0.86)
- EfficientNav: Towards On-Device Object-Goal Navigation with Navigation Map Caching and Retrieval (NeurIPS 2025 · 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.
