AstraNav-Memory：面向长时记忆的上下文压缩
AstraNav-Memory: Contexts Compression for Long Memory
用冻结 DINOv3 特征加 PixelUnshuffle+Conv 构成的轻量视觉分词器把每帧压到约 30 个 token，使 VLM 策略能在单上下文里塞进数百帧历史；这套「图像中心压缩记忆」可直接迁移到任何需要长视觉历史输入的 VLM/具身策略中。
原文依据 our visual tokenizer reduces native vision tokens by roughly 10-20x, representing each image with about 30 tokens
摘要译文
由模型翻译 · 术语统一 · 请对照原文核对关键结论终身具身导航要求智能体跨任务积累、保持并利用空间-语义经验，从而在新环境中高效探索、在熟悉环境中快速到达目标。以物体为中心的记忆虽具可解释性，但依赖检测与重建流程，限制了鲁棒性与可扩展性。我们提出一种以图像为中心的记忆框架，通过高效的视觉上下文压缩模块与基于 Qwen2.5-VL 的导航策略端到端耦合，实现长期隐式记忆。我们的视觉分词器构建于冻结 DINOv3 特征的 ViT 骨干网络之上，并采用轻量 PixelUnshuffle+Conv 模块，将原生视觉 token 减少约 10-20 倍，每张图像仅用约 30 个 token 表示，使智能体能在单个上下文内保持数百帧历史。在 GOAT-Bench 与 HM3D-OVON 上的实验结果表明，本方法取得最优的导航性能，改善了陌生环境中的探索，并缩短了熟悉环境中的路径。消融研究进一步表明，适度压缩在效率与准确率之间取得最佳平衡。这些发现使压缩的以图像为中心的记忆成为终身具身智能体一种实用且可扩展的接口，使其能够对长视觉历史进行推理，并以类人的效率导航。
英文原文摘要 1399 字符
Lifelong embodied navigation requires agents to accumulate, retain, and exploit spatial-semantic experience across tasks, enabling efficient exploration in novel environments and rapid goal reaching in familiar ones. While object-centric memory is interpretable, it depends on detection and reconstruction pipelines that limit robustness and scalability. We propose an image-centric memory framework that achieves long-term implicit memory via an efficient visual context compression module end-to-end coupled with a Qwen2.5-VL-based navigation policy. Built atop a ViT backbone with frozen DINOv3 features and lightweight PixelUnshuffle+Conv blocks, our visual tokenizer reduces native vision tokens by roughly 10-20x, representing each image with about 30 tokens and allowing the agent to maintain hundreds of historical frames within a single context. Experimental results on GOAT-Bench and HM3D-OVON show that our method achieves state-of-the-art navigation performance, improving exploration in unfamiliar environments and shortening paths in familiar ones. Ablation studies further reveal that moderate compression provides the best balance between efficiency and accuracy. These findings position compressed image-centric memory as a practical and scalable interface for lifelong embodied agents, enabling them to reason over long visual histories and navigate with human-inspired efficiency.