AstraNav-Memory is part of the Astra series for robust embodied intelligence:
- AstraNav-Memory (This Repo): Focuses on long-term implicit memory through visual context compression.
- AstraNav-World: Focuses on predictive planning via tightly coupled visual foresight and action generation.
- OmniNav: Focuses on unified multi-paradigm navigation and real-time execution via a fast-slow system.
🤏 20× Visual Context Compression: Employs a lightweight PixelUnshuffle+Conv tokenizer to compress frames into ~30 tokens, expanding context capacity from tens to hundreds of images for massive, long-term implicit memory.
🎞️ Qwen-DINO Unified Policy: Couples Qwen2.5-VL reasoning with frozen DINOv3 features in an end-to-end framework, replacing fragile object-centric pipelines with a robust, scalable image-centric memory interface.
🗺️ Lifelong Navigational Mastery: Sets new SOTA benchmarks on GOAT-Bench and HM3D-OVON by balancing efficient exploration in novel environments with optimal, high-speed pathfinding in familiar ones.
- January 20, 2026: We release the inference checkpoints with various downsampling rates and context lengths.
- December 30, 2025: We release the training and inference code.
Clone the repo:
git clone https://github.com/amap-cvlab/AstraNav-Memory.git
Install Training dependencies:
# Ensure torch >= 2.6.0
cd train_code
pip install -r requirements.txt
Install habitat-sim and habitat-lab for inference
● habitat-sim
git clone https://github.com/facebookresearch/habitat-sim.git && cd habitat-sim && git checkout v0.2.3
pip install -r requirements.txt
python setup.py install --headless
● habitat-lab
git clone https://github.com/chongchong2025/habitat-lab && cd habitat-lab && git checkout v0.2.3_waypoint
python -m pip install -r habitat-baselines/habitat_baselines/rl/requirements.txt
python -m pip install -r habitat-baselines/habitat_baselines/rl/ddppo/requirements.txt
pip install -e .
cd habitat-baselines
pip install -e .
| Models | Download Link |
|---|---|
| DS16-Context50 | ModelScope |
| DS16-Context100 | ModelScope |
| DS4-Context100 | ModelScope |
| DS64-Context100 | ModelScope |
● Goat-Bench
cd inference_code/hm3d-online
python goat-nav.py● OVON
cd inference_code/hm3d-online
python ovon-nav.pycd train_code
bash run_train.shIf you find this repository useful, please consider giving a star ⭐ and citation
@article{ren2025astranav-memory,
title={AstraNav-Memory: Contexts Compression for Long Memory},
author={Botao Ren and Junjun Hu and Xinda Xue and Minghua Luo and Jintao Chen and Haochen Bai and Liangliang You and Mu Xu},
year={2025},
eprint={2512.21627},
}
Thanks to OmniNav, MTU3D, and OVON for open-sourcing the construction of training data and the closed-loop inference code. Their contributions have significantly enriched the open-source community.