[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"doc-detail-82794-en":3,"doc-seo-82794-105":29,"detail-sidebar-cat-0-en-105":91},{"code":4,"msg":5,"data":6},0,"success",{"doc_id":7,"user_id":8,"nickname":9,"user_avatar":10,"doc_module":4,"category_id":11,"category_name":12,"doc_title":13,"doc_description":14,"doc_content":15,"file_id":16,"file_url":17,"file_type":18,"file_size":19,"view_count":20,"is_deleted":4,"is_public":20,"is_downloadable":20,"audit_status":20,"page_count":21,"language":22,"language_code":23,"site_id":24,"html_lang":23,"table_of_contents":25,"faqs":26,"seo_title":13,"seo_description":14,"update_tm":27,"read_time":28},82794,137441390410,"Hazel","https://ap-avatar.wpscdn.com/avatar/2000252f4ab5702993?_k=1776741390130283984",8,"Research & Report","XS-VLA Coupling Coarse-grained Spatial Distillation with Latent Flow Matching for Lightweight Robotic Control","Large Vision-Language Models provide strong spatial grounding for multimodal understanding but are computationally prohibitive for real-time robotic control, while lightweight models enable edge deployment yet suffer from “spatial blindness” in spatial prediction. Training Vision-Language-Action policies on heterogeneous human demonstrations can further degrade performance. XS-VLA introduces a two-stage framework: distill coarse-grained spatial semantics into a SmolVLM2-0.25B backbone, then couple a CVAE with flow matching dynamics for robust multimodal control. Experiments on LIBERO show new SOTA results for models under 0.5B, with up to 7.2% average success-rate gains.","XS-VLA: Coupling Coarse-grained Spatial Distillation with Latent Flow Matching for Lightweight Robotic Control  \nLei Iok Tong 1 , Qingchen Xie 1 , Wei Huang 1 , Ying Jie Yap 1 , Yujie Zhang2 , Qianzhi Li3 , Xiaolong Liu* ,  \nZhidong Deng 1,*  \narXiv :2607 .04 17 1v 1 [ cs .RO] 5 Jul 2026  \nAbstract—While Large Vision-Language Models (LVLMs) have revolutionized multimodal understanding with strong spatial grounding, they remain computationally prohibitive for real-time robotic control. Conversely, lightweight models offer the efficiency required for edge deployment but suffer from “spatial blindness”—a lack or weakness of native spatial prediction capabilities. Furthermore, training Vision-LanguageAction (VLA) models on mixed human demonstration data often degrades policy performance due to the complexity of modeling highly variant behaviors. These limitations severely hamper the utility of lightweight models in precise robotic manipulation. To bridge this gap, we propose XS-VLA, a novel two-stage framework. First, we distill spatial semantic knowledge from a capable teacher model (Qwen3-VL- 4B) into the SmolVLM2-0.25B backbone. By fine-tuning on a curated dataset of coarse-grained spatial descriptions, we transform this lightweight model into a spatially grounded engine. Second, we leverage this enhanced backbone to condition a novel Latent Flow Matching policy. Unlike standard deterministic controllers, our policy couples a Conditional Variational Autoencoder (CVAE) with Flow Matching dynamics to robustly model complex, multimodal action distributions. Our model achieves new SOTA performance among models \u003C0.5B on the Libero benchmark. Extensive experiments on the LIBERO benchmark demonstrate that this “distill-thencontrol” approach significantly enhances performance. XS-VLA achieves up to a 7.2% increase in success rates on average (23% increase on Libero-Long) compared to the SmolVLA 0.25B baseline and even outperforms the significantly larger 2.2B vanilla SmolVLA. Ultimately, in our ablation study, our framework demonstrates that specialized spatial tuning combined with generative latent flow control dramatically improves lightweight VLA performance, delivering a 3.2 × speedup in mission execution compared to the previous lightweight flow matching policy.  \nI. INTRODUCTION  \nThe field of Embodied AI has long sought to create general-purpose robots capable of understanding unstructured environments and executing complex instructions. The recent emergence of Large Vision-Language Models (LVLMs), such as GPT-4V [1], Gemini, and Qwen3-VL [2], has brought us closer to this goal. These models demonstrate  \n1 Iok Tong Lei, Qingchen Xie, Wei Huang, Ying Jie Yap, and Zhidong Deng are with the Department of Computer Science and Technology, Tsinghua University, Beijing, China.  \n2 Qianzhi Li is with the National College for Excellent Engineers, Beihang University, Beijing, China.  \n3 Xiaolong Liu is with the Wuxi Dexteroushands Robotic Technology Co.  \n4 Yujie Zhang is an independent researcher.  \n* Corresponding authors: Xiaolong Liu and Zhidong Deng. Emails: [xllau@126.com](xllau@126.com), [michael@mail.tsinghua.edu.cn](michael@mail.tsinghua.edu.cn. Primary)[. Primary](michael@mail.tsinghua.edu.cn. Primary)[ ](michael@mail.tsinghua.edu.cn. Primary)contact: Iok Tong Lei, [lixt22@mails.tsinghua.edu.cn](lixt22@mails.tsinghua.edu.cn).  \nan unprecedented ability to interpret visual data, reason about object affordances, and generate high-level plans. For instance, an LVLM can look at a messy kitchen counter and not only identify a “spilled cup” but also reason that it needs to be picked up and placed in the sink.  \nHowever, a critical bottleneck remains in the deployment of these models: the trade-off between reasoning capability and inference latency. Robotic control loops, particularly for manipulation tasks involving contact, typically require control frequencies between 10Hz and 50Hz to remain stable and reactive. LVLMs, with paramet","cbCaikCjTEmBbl7H","https://ap.wps.com/l/cbCaikCjTEmBbl7H","pdf",2427397,1,9,"English","en",105,"# Introduction\n## Embodied AI and vision-language models\n## Deployment bottleneck: latency vs reasoning\n## Spatial grounding gap in lightweight models\n## Proposed approach: XS-VLA two-stage framework","[{\"question\":\"Why are large vision-language models difficult to use for real-time robotic control?\",\"answer\":\"They typically have very large parameter counts and run at inference frequencies far below the 10–50Hz needed for stable, reactive manipulation control loops.\"},{\"question\":\"What is “spatial blindness” in lightweight models?\",\"answer\":\"Lightweight VLMs are often not trained with extensive spatial description data, so they recognize objects without accurately encoding spatial coordinates and relations required for grasping.\"},{\"question\":\"How does XS-VLA improve lightweight robotic control performance?\",\"answer\":\"XS-VLA first distills coarse-grained spatial semantic knowledge into a 0.25B backbone, then uses a Latent Flow Matching policy that couples a CVAE with flow matching dynamics to model complex multimodal action distributions.\"}]",1784182973,23,{"code":4,"msg":30,"data":31},"ok",{"site_id":24,"language":23,"slug":32,"title":13,"keywords":33,"description":14,"schema_data":34,"social_meta":86,"head_meta":88,"extra_data":90,"updated_unix":27},"xs-vla-coupling-coarse-grained-spatial-distillation-with-latent-flow-matching-for-lightweight-robotic-control","",{"@graph":35,"@context":85},[36,53,68],{"@type":37,"itemListElement":38},"BreadcrumbList",[39,43,47,50],{"item":40,"name":41,"@type":42,"position":20},"https://docshare.wps.com","Home","ListItem",{"item":44,"name":45,"@type":42,"position":46},"https://docshare.wps.com/document/","Document",2,{"item":48,"name":12,"@type":42,"position":49},"https://docshare.wps.com/document/research-report/",3,{"item":51,"name":13,"@type":42,"position":52},"https://docshare.wps.com/document/xs-vla-coupling-coarse-grained-spatial-distillation-with-latent-flow-matching-for-lightweight-robotic-control/82794/",4,{"url":51,"name":13,"@type":54,"author":55,"headline":13,"publisher":57,"fileFormat":60,"inLanguage":23,"description":14,"dateModified":61,"datePublished":62,"encodingFormat":60,"isAccessibleForFree":63,"interactionStatistic":64},"DigitalDocument",{"name":9,"@type":56},"Person",{"url":40,"name":58,"@type":59},"DocShare","Organization","application/pdf","2026-07-17","2026-07-16",true,{"@type":65,"interactionType":66,"userInteractionCount":20},"InteractionCounter",{"@type":67},"ViewAction",{"@type":69,"mainEntity":70},"FAQPage",[71,77,81],{"name":72,"@type":73,"acceptedAnswer":74},"Why are large vision-language models difficult to use for real-time robotic control?","Question",{"text":75,"@type":76},"They typically have very large parameter counts and run at inference frequencies far below the 10–50Hz needed for stable, reactive manipulation control loops.","Answer",{"name":78,"@type":73,"acceptedAnswer":79},"What is “spatial blindness” in lightweight models?",{"text":80,"@type":76},"Lightweight VLMs are often not trained with extensive spatial description data, so they recognize objects without accurately encoding spatial coordinates and relations required for grasping.",{"name":82,"@type":73,"acceptedAnswer":83},"How does XS-VLA improve lightweight robotic control performance?",{"text":84,"@type":76},"XS-VLA first distills coarse-grained spatial semantic knowledge into a 0.25B backbone, then uses a Latent Flow Matching policy that couples a CVAE with flow matching dynamics to model complex multimodal action distributions.","https://schema.org",{"og:url":51,"og:type":87,"og:title":13,"og:site_name":58,"og:description":14},"article",{"robots":89,"canonical":51},"index,follow",{"doc_id":7,"site_id":24},{"code":4,"msg":5,"data":92},[93,97,101,105,110,115,120,123,127,130,134],{"id":20,"doc_module":4,"doc_module_name":45,"category_name":94,"show_sort_weight":95,"slug":96},"Story & Novel",90,"story-novel",{"id":46,"doc_module":4,"doc_module_name":45,"category_name":98,"show_sort_weight":99,"slug":100},"Literature",80,"literature",{"id":52,"doc_module":4,"doc_module_name":45,"category_name":102,"show_sort_weight":103,"slug":104},"Exam",70,"exam",{"id":106,"doc_module":4,"doc_module_name":45,"category_name":107,"show_sort_weight":108,"slug":109},5,"Comic",60,"comic",{"id":111,"doc_module":4,"doc_module_name":45,"category_name":112,"show_sort_weight":113,"slug":114},6,"Technology",50,"technology",{"id":116,"doc_module":4,"doc_module_name":45,"category_name":117,"show_sort_weight":118,"slug":119},7,"Healthcare",40,"healthcare",{"id":11,"doc_module":4,"doc_module_name":45,"category_name":12,"show_sort_weight":121,"slug":122},30,"research-report",{"id":21,"doc_module":4,"doc_module_name":45,"category_name":124,"show_sort_weight":125,"slug":126},"Religion & Spirituality",20,"religion-spirituality",{"id":125,"doc_module":4,"doc_module_name":45,"category_name":128,"show_sort_weight":125,"slug":129},"World Cup","world-cup",{"id":131,"doc_module":4,"doc_module_name":45,"category_name":132,"show_sort_weight":131,"slug":133},10,"Lifestyle","lifestyle",{"id":135,"doc_module":4,"doc_module_name":45,"category_name":136,"show_sort_weight":106,"slug":137},19,"General","general"]