[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"doc-detail-83819-en":3,"doc-seo-83819-105":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},83819,5909877438554,"Maeve","https://ap-avatar.wpscdn.com/avatar/5600025385ad2bf12a7?_k=1778553567797529272",8,"Research & Report","VLA Grounder Language Conditioning Space Optimization for Black Box VLA Models","Vision-Language-Action (VLA) models are treated as end-to-end action policies conditioned on natural-language tasks, yet performance can vary sharply with instruction phrasing. This work studies whether frozen VLA policies can be improved without updating action weights by optimizing a dedicated language-conditioning space. A language-conditioning space policy converts a human instruction into a short VLA-grounded command using visual object appearance, spatial relations, and target-grounding cues, trained with RL from sparse success rewards. Experiments on RL4VLA and VL-Think show improved success across instruction-sensitive, symbolic, and multi-object manipulation tasks, highlighting language as an optimizable robot foundation-model variable.","VLA Grounder: Language-Conditioning Space Optimization for Black-Box  \nVLA Models  \nDamir Shodiev2 , Aleksei Staroverov 1 ,2 ,3 , Nikita Kachaev 1 , Alexey K. Kovalev 1 ,2 , Aleksandr I. Panov 1 ,2  \n1AXXX, 2MIRAI, 3MISIS  \n[damir.shodiev.pro@gmail.com](damir.shodiev.pro@gmail.com)  \narXiv :2607 .045 17v 1 [ cs .AI ] 5 Jul 2026  \nAbstract  \nVision-Language-Action (VLA) models are commonly treated as end-to-end action policies conditioned on natural-language task descriptions. In practice, however, their behavior often depends sharply on how the instruction is phrased, suggesting that language is not merely a task label but an optimizable conditioning input. We study whether frozen VLA policies can be improved by optimizing language space rather than updating action weights. Our method introduces a language-conditioning space policy that translates a human instruction into a short VLAgrounded command using object appearance, spatial relations, and target-grounding cues. The language-conditioning space policy is initialized with a failure-derived command-space prior and optimized with reinforcement learning from sparse task-completion rewards, while the downstream VLA remains fully frozen. This yields language-conditioning space optimization: RL discovers which VLA-grounded commands best elicit successful behavior from the frozen action policy. Experiments on RL4VLA and VL-Think show that languageconditioning space optimization improves success on instruction-sensitive, symbolic, and multi-object manipulation tasks, demonstrating that language can serve as an optimizable variable for a robot foundation models.  \nWebsite: [https://tttonyalpha.github](https://tttonyalpha.github).  \nio/vla_grounder  \n1 Introduction  \nVision-Language-Action (VLA) models represent a promising paradigm in robotic learning, combining visual perception, natural language understanding, and action generation within a single policy. Recent VLA models leverage large-scale pre-training and robotic demonstrations to produce actions directly from images and language, enabling broad  \nFigure 1: Visual overview of language-conditioning space optimization. A failure-derived command-space prior conditions a pre-trained VLM to translate the user instruction and scene image into a short VLA-grounded command. This command conditions a VLA policy, improving action success without updating VLA weights.  \nzero-shot manipulation across tasks and embodiments (Black et al., 2024 ; Zhang et al., 2025 ; Zhong et al., 2025) .  \nDespite this progress, the language input to a VLA is often treated as a passive task description, even though it directly shapes the action distribution of the policy. In practice, VLA behavior is highly sensitive to how a goal is expressed (Pugacheva et al., 2025 ; Kachaev et al., 2025) . A command such as \"put bread on plate\" may fail even when the scene contains a visually clear target, while a more grounded command such as \"pick up the brown round object and put it on the yellow  \nplate\" can succeed. Similar effects appear when an object name is visually misleading: \"put champagne glass on plate\" can be improved by referring to the same object as \"the tall white glass\". These examples suggest that many VLA failures arise from an instruction-to-grounding mismatch between human semantic intent and the perceptual categories available to the robot policy.  \nThis mismatch is amplified by the limited linguistic diversity of robot demonstration datasets (Walke et al., 2024 ; Collaboration et al., 2025) . Most datasets contain short templated instructions and everyday household objects, while benchmark tasks may include ambiguous objects, abstract references, uncommon categories, or multiple visually similar distractors. As a result, the language channel of a VLA is underused: models reserve substantial capacity for text, but the instruction often fails to provide the visual and spatial cues needed for robust action selection (Zhang et al., 2025 ; Liu et ","cbCaiev5UEjy5KTm","https://ap.wps.com/l/cbCaiev5UEjy5KTm","pdf",4520352,1,14,"English","en",105,"# Abstract\n# Introduction\n## Language sensitivity in VLA instruction phrasing\n## Mismatch between human intent and robot grounding categories\n## Data-driven underuse of the language channel\n## Proposed language-conditioning space policy (VLA Grounder)","[{\"question\":\"What problem does VLA Grounder address in black-box VLA models?\",\"answer\":\"It addresses the observation that VLA behavior depends strongly on how instructions are phrased, indicating that language is more than a passive label. Failures can occur when human semantic intent does not align with the robot’s perceptual grounding categories.\"},{\"question\":\"How does the method improve a frozen VLA without changing its action weights?\",\"answer\":\"It inserts a language-conditioning space policy upstream of the frozen VLA. The policy rewrites the instruction into a short VLA-grounded command using object appearance, spatial relations, and target-grounding cues, adapting the conditioning signal while keeping the downstream model fixed.\"},{\"question\":\"How is the language-conditioning space policy trained?\",\"answer\":\"It is trained with reinforcement learning from sparse task-completion rewards, while the downstream VLA remains fully frozen. A failure-derived command-space prior guides exploration toward concise command forms.\"}]",1784190621,35,{"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},"vla-grounder-language-conditioning-space-optimization-for-black-box-vla-models","",{"@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/vla-grounder-language-conditioning-space-optimization-for-black-box-vla-models/83819/",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},"What problem does VLA Grounder address in black-box VLA models?","Question",{"text":75,"@type":76},"It addresses the observation that VLA behavior depends strongly on how instructions are phrased, indicating that language is more than a passive label. Failures can occur when human semantic intent does not align with the robot’s perceptual grounding categories.","Answer",{"name":78,"@type":73,"acceptedAnswer":79},"How does the method improve a frozen VLA without changing its action weights?",{"text":80,"@type":76},"It inserts a language-conditioning space policy upstream of the frozen VLA. The policy rewrites the instruction into a short VLA-grounded command using object appearance, spatial relations, and target-grounding cues, adapting the conditioning signal while keeping the downstream model fixed.",{"name":82,"@type":73,"acceptedAnswer":83},"How is the language-conditioning space policy trained?",{"text":84,"@type":76},"It is trained with reinforcement learning from sparse task-completion rewards, while the downstream VLA remains fully frozen. A failure-derived command-space prior guides exploration toward concise command forms.","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},[93,97,101,105,110,115,120,123,128,131,135],{"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":124,"doc_module":4,"doc_module_name":45,"category_name":125,"show_sort_weight":126,"slug":127},9,"Religion & Spirituality",20,"religion-spirituality",{"id":126,"doc_module":4,"doc_module_name":45,"category_name":129,"show_sort_weight":126,"slug":130},"World Cup","world-cup",{"id":132,"doc_module":4,"doc_module_name":45,"category_name":133,"show_sort_weight":132,"slug":134},10,"Lifestyle","lifestyle",{"id":136,"doc_module":4,"doc_module_name":45,"category_name":137,"show_sort_weight":106,"slug":138},19,"General","general"]