arXiv cs.AI· Xinyu Liu, Daiying Tian, Chenwei Tang, Ruyang Han, Chengyang Zhang, Qingyu Zeng, Wei Ju, Liangli Zhen, Jiancheng Lv·· 12 小时前AI 评分34
CogMemory：面向长程 VLA 模型的变分认知记忆抽象框架
Variational Cognitive Memory Abstraction for Long-Horizon Vision-Language-Action Models
AI 导读
研究者提出 CogMemory，一个变分认知记忆框架，让 VLA 系统能选择性保留并复用任务相关的历史信息。该方法通过选择性状态空间记忆编码器、多槽时间抽象与变分信息压缩，将多模态轨迹编码为紧凑的潜在记忆表示，并在动作生成时用动态交叉注意力检索与门控融合把相关记忆注入策略预测。仅增加 3.45M 参数，即可适配早期融合与中期融合两类 VLA 主干，在长程机器人操作基准上取得一致提升。
正文
Abstract:Vision-Language-Action (VLA) models have shown remarkable success in robotic manipulation, yet their ability to execute long-horizon tasks remains limited by insufficient long- term memory modeling. Existing methods either rely on implicit temporal representations within policy networks, where historical information can be gradually diluted, or maintain explicit memory buffers that suffer from redundancy, noise accumulation, and inefficient retrieval. In this paper, we propose CogMemory, a novel variational cognitive memory framework that enables VLA systems to selectively preserve and reuse task-relevant historical information. Our approach encodes multimodal trajectories into compact latent memory representations through a selective state- space memory encoder, multi-slot temporal abstraction, and variational information compression. During action generation, dynamic cross-attention retrieval with gated fusion adaptively incorporates relevant memories into policy prediction. With merely 3.45M additional parameters, CogMemory stays lightweight and seamlessly fits both early- and intermediate-fusion VLA backbones. Extensive experiments on long-horizon robotic manipulation benchmarks demonstrate consistent improvements across diverse VLA backbones.
| Subjects: | Artificial Intelligence (cs.AI) |
| Cite as: | arXiv:2610.34251 [cs.AI] |
| (or arXiv:2610.34251v1 [cs.AI] for this version) | |
| https://doi.org/10.48550/arXiv.2610.34251 arXiv-issued DOI via DataCite (pending registration) |
Submission history
From: Xinyu Liu [view email]
[v1]
Sun, 25 Oct 2026 12:24:57 UTC (1,511 KB)
来源：arXiv cs.AI · arxiv.org