[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"detail-sidebar-cat-0-en-105":3,"doc-seo-327505-105":59,"doc-detail-327505-en":130},{"code":4,"msg":5,"data":6},0,"success",[7,13,18,23,28,33,38,43,48,51,55],{"id":8,"doc_module":4,"doc_module_name":9,"category_name":10,"show_sort_weight":11,"slug":12},1,"Document","Story & Novel",90,"story-novel",{"id":14,"doc_module":4,"doc_module_name":9,"category_name":15,"show_sort_weight":16,"slug":17},2,"Literature",80,"literature",{"id":19,"doc_module":4,"doc_module_name":9,"category_name":20,"show_sort_weight":21,"slug":22},4,"Exam",70,"exam",{"id":24,"doc_module":4,"doc_module_name":9,"category_name":25,"show_sort_weight":26,"slug":27},5,"Comic",60,"comic",{"id":29,"doc_module":4,"doc_module_name":9,"category_name":30,"show_sort_weight":31,"slug":32},6,"Technology",50,"technology",{"id":34,"doc_module":4,"doc_module_name":9,"category_name":35,"show_sort_weight":36,"slug":37},7,"Healthcare",40,"healthcare",{"id":39,"doc_module":4,"doc_module_name":9,"category_name":40,"show_sort_weight":41,"slug":42},8,"Research & Report",30,"research-report",{"id":44,"doc_module":4,"doc_module_name":9,"category_name":45,"show_sort_weight":46,"slug":47},9,"Religion & Spirituality",20,"religion-spirituality",{"id":46,"doc_module":4,"doc_module_name":9,"category_name":49,"show_sort_weight":46,"slug":50},"World Cup","world-cup",{"id":52,"doc_module":4,"doc_module_name":9,"category_name":53,"show_sort_weight":52,"slug":54},10,"Lifestyle","lifestyle",{"id":56,"doc_module":4,"doc_module_name":9,"category_name":57,"show_sort_weight":24,"slug":58},19,"General","general",{"code":4,"msg":60,"data":61},"ok",{"site_id":62,"language":63,"slug":64,"title":65,"keywords":66,"description":67,"schema_data":68,"social_meta":123,"head_meta":125,"extra_data":127,"updated_unix":129},105,"en","learning-to-move-before-learning-to-do-task-agnostic-pretraining-for-vlas","Learning to Move Before Learning to Do - Task-Agnostic pretraining for VLAs","","Vision-Language-Action (VLA) models struggle with the limited availability of expert demonstrations consisting of observation, instruction, and action triplets. The work attributes this bottleneck to mixing two different learning goals: acquiring physical competence (“how to move”) and semantic alignment (“what to do”), where only the latter truly needs language supervision. It introduces Task-Agnostic Pretraining (TAP), a two-stage method that learns transferable motor priors from unlabeled interaction data using self-supervised inverse dynamics, then grounds them in language with minimal expert data.",{"@graph":69,"@context":122},[70,84,105],{"@type":71,"itemListElement":72},"BreadcrumbList",[73,77,79,82],{"item":74,"name":75,"@type":76,"position":8},"https://docshare.wps.com","Home","ListItem",{"item":78,"name":9,"@type":76,"position":14},"https://docshare.wps.com/document/",{"item":80,"name":40,"@type":76,"position":81},"https://docshare.wps.com/document/research-report/",3,{"item":83,"name":65,"@type":76,"position":19},"https://docshare.wps.com/document/learning-to-move-before-learning-to-do-task-agnostic-pretraining-for-vlas/327505/",{"url":83,"name":65,"@type":85,"image":86,"author":91,"headline":65,"publisher":94,"fileFormat":97,"inLanguage":63,"description":67,"dateModified":98,"datePublished":99,"encodingFormat":97,"isAccessibleForFree":100,"interactionStatistic":101},"DigitalDocument",{"url":87,"@type":88,"width":89,"height":90},"https://docshare.wps.com/thumbnails/learning-to-move-before-learning-to-do-task-agnostic-pretraining-for-vlas/327505.png","ImageObject",300,407,{"name":92,"@type":93},"Hazel","Person",{"url":74,"name":95,"@type":96},"DocShare","Organization","application/pdf","2026-10-24","2026-09-21",true,{"@type":102,"interactionType":103,"userInteractionCount":52},"InteractionCounter",{"@type":104},"ViewAction",{"@type":106,"mainEntity":107},"FAQPage",[108,114,118],{"name":109,"@type":110,"acceptedAnswer":111},"Why are expert demonstrations a bottleneck for VLA models?","Question",{"text":112,"@type":113},"VLA training relies on costly-to-collect expert triplets of observations, language instructions, and actions. The scarcity and labor cost of collecting them at scale limits performance and generality.","Answer",{"name":115,"@type":110,"acceptedAnswer":116},"What is the Decomposition Hypothesis proposed in the work?",{"text":117,"@type":113},"It separates two learning objectives: physical competence (“how to move”) and semantic alignment (“what to do”). The key claim is that only semantic alignment requires language supervision.",{"name":119,"@type":110,"acceptedAnswer":120},"How does Task-Agnostic Pretraining (TAP) use unlabeled data?",{"text":121,"@type":113},"TAP first learns transferable motor priors from cheap unlabeled interaction data by training with a self-supervised Inverse Dynamics objective. A lightweight second stage then uses minimal expert data to connect those priors to language instructions.","https://schema.org",{"og:url":83,"og:type":124,"og:title":65,"og:site_name":95,"og:description":67},"article",{"robots":126,"canonical":83},"index,follow",{"doc_id":128,"site_id":62},327505,1790423054,{"code":4,"msg":5,"data":131},{"doc_id":128,"user_id":132,"nickname":92,"user_avatar":133,"doc_module":4,"category_id":39,"category_name":40,"doc_title":65,"doc_description":67,"doc_content":134,"file_id":135,"file_url":136,"file_type":137,"file_size":138,"view_count":52,"is_deleted":4,"is_public":8,"is_downloadable":8,"audit_status":8,"page_count":139,"language":140,"language_code":63,"site_id":62,"html_lang":63,"table_of_contents":141,"faqs":142,"seo_title":143,"seo_description":67,"update_tm":144,"read_time":145,"preview_page_count":4},137441390410,"https://ap-avatar.wpscdn.com/avatar/2000252f4ab5702993?_k=1776741390130283984","Learning to Move Before Learning to Do: Task-Agnostic pretraining for VLAs  \nJunhao Shi 1,2 Siyin Wang 1,2 Xiaopeng Yu 1 Li Ji 1 Jingjing Gong2,† Xipeng Qiu 1,2,†  \n[24110240071@m.fudan.edu.cn](24110240071@m.fudan.edu.cn)  \n1 Fudan University 2 Shanghai Innovation Institute  \narXiv :2607 .02466v 1 [ cs .RO] 2 Jul 2026  \nAbstract  \nVision-Language-Action (VLA) models are fundamentally bottlenecked by the scarcity of expert demonstrations—triplets of observations, instructions, and actions that are costly to collect at scale.  \nWe argue that this bottleneck stems from conflating two distinct learning objectives: acquiring physical competence (how to move) and acquiring semantic alignment (what to do) . Crucially, only the latter requires language supervision. Building on this Decomposition Hypothesis, we propose Task-Agnostic Pretraining (TAP), a two-stage framework that first learns transferable motor priors from cheap, unlabeled interaction data—including discarded off-task trajectories and autonomous robot play—via a self-supervised Inverse Dynamics objective. A lightweight second stage then grounds these priors in language using minimal expert data. On the SIMPLER benchmark, TAP matches models trained on over 1M expert trajectories while using orders of magnitude less labeled data, yielding a 10% absolute gain over standard behavior cloning. On a real-world WidowX platform, TAP retains 25% success under camera perturbations where internet-scale baselines collapse to 0%, demonstrating that taskagnostic pretraining produces robust, transferable physical representations and offers a scalable path forward for Embodied AI.  \nHomepage: [https://sjh0354.github.io/task_agnostic_pretrain](https://sjh0354.github.io/task_agnostic_pretrain)  \nGithub Repo: [https://github.com/sjh0354/Task-Agnostic-Pretrain](https://github.com/sjh0354/Task-Agnostic-Pretrain)  \nHF Models: [https://huggingface.co/collections/Michael0354/task-agnostic-pretrain](https://huggingface.co/collections/Michael0354/task-agnostic-pretrain)  \n1 Introduction  \nThe development of Vision-Language-Action (VLA) models [1–5] has opened promising avenues for building general-purpose robots. However, training these models requires an enormous amount of high-quality robotic data. The predominant approach to acquiring such data relies on human teleoperation, where expert operators guide robot movements while providing language based task annotations [6–8] . This paradigm is inherently passive from the robot’s perspective: the robot serves merely as a vessel for capturing human demonstrations, contributing no exploration of its own. Beyond being prohibitively expensive and laborintensive, this data collection scheme is fundamentally unnatural—it conflates the embodied experience of a  \nrobot with the disembodied intentions of a human operator. Crucially, it is also non-scalable: the rate of data acquisition is bottlenecked by the availability and endurance of human operators, making it impractical to  \n† Corresponding Authors.  \ncollect the diverse, large-scale datasets required for truly general-purpose manipulation.  \nConsider, by contrast, how biological agents acquire motor competence. A human infant does not learn to grasp, manipulate, and interact with objects by passively receiving demonstrations from an expert. Instead, infants engage in spontaneous, task-agnostic exploration—reaching, touching, dropping, and observing the consequences of their actions [9] . This curiosity-driven self-exploration is philosophically task-unaware: the infant is not optimizing for any specific goal but is rather building an internal model of how the world responds to its actions [10] . Through this process, the infant develops a rich understanding of physics, affordances, and sensorimotor contingencies [11] long before any explicit task instruction is given. The grounding in “how to move” emerges naturally from active interaction, decoupled from “what to do.”  \nInspired by this developmenta","cbCaigKpzCu3FiRm","https://ap.wps.com/l/cbCaigKpzCu3FiRm","pdf",5519740,21,"English","# Abstract\n# Introduction","[{\"question\":\"Why are expert demonstrations a bottleneck for VLA models?\",\"answer\":\"VLA training relies on costly-to-collect expert triplets of observations, language instructions, and actions. The scarcity and labor cost of collecting them at scale limits performance and generality.\"},{\"question\":\"What is the Decomposition Hypothesis proposed in the work?\",\"answer\":\"It separates two learning objectives: physical competence (“how to move”) and semantic alignment (“what to do”). The key claim is that only semantic alignment requires language supervision.\"},{\"question\":\"How does Task-Agnostic Pretraining (TAP) use unlabeled data?\",\"answer\":\"TAP first learns transferable motor priors from cheap unlabeled interaction data by training with a self-supervised Inverse Dynamics objective. A lightweight second stage then uses minimal expert data to connect those priors to language instructions.\"}]","Learning to Move Before Learning to Do - Task-Agnostic pretraining for VLAs | PDF",1789977367,53]