[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"doc-detail-82581-en":3,"doc-seo-82581-105":29,"detail-sidebar-cat-0-en-105":90},{"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":4,"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},82581,34359740700684,"Finn","https://ap-avatar.wpscdn.com/avatar/1f400023980c374ae676?_k=1777273430885731487",8,"Research & Report","ROSA Robotics Foundation Model Serving System for Robot Factories","Robotics foundation models are increasingly viable for factory deployments, but existing serving systems assume a single robot and single model, optimizing mainly single-action latency. ROSA (Robotics Oriented Serving Architecture) redesigns serving for robot factories with three principles: shared GPU-pool serving over the network, robotics-aware abstractions for multi-model pipelines and failure handling, and scheduling driven by factory productivity SLOs. Implemented on Ray Serve with vLLM, PyTorch, and JAX, ROSA improves factory productivity by up to 12.06× versus dedicated systems.","ROSA: A Robotics Foundation Model Serving System for Robot Factories  \nWenqi Jiang1 Jason Clemons1 Rowland O’Flaherty1 Hugo Hadfield1 AlperenDegirmenci1 Shuran Song1,2 Yashraj Narang1 Christos Kozyrakis1,2  \n1NVIDIA Research 2 Stanford University  \narXiv :2607 .0 1088v 1 [ cs .RO] 1 Jul 2026  \nAbstract  \nRobotics foundation models (RFMs) are making generalpurpose robots increasingly practical for factory deployments. While RFM serving systems are central to this vision, existing systems are largely shaped by a single-robot, single-model assumption: inference is treated as an edge-computing problem handled by an on-robot or dedicated nearby GPU, and the serving objective is to minimizing the latency of a single action model. In this paper, we propose ROSA, the Robotics Oriented Serving Architecture, an RFM serving system for robot factories designed around three key principles. First, ROSA adopts shared GPU-pool serving, allowing a fleet of robots to access powerful server-class GPUs over the network in order to improve inference performance, battery duration, and GPU utilization. Second, ROSA provides a robotics-aware programming abstraction and system design that supports multi-model pipelines, per-task performance requirements, and failure handling. Third, ROSA usesfactory-objective-driven scheduling to maximize SLO-qualified factory productivity rather than minimizing individual request latency. We implement ROSA on top ofRay Serve for distributed orchestration, with vLLM, PyTorch, and JAX as model-serving backends, and evaluate it on both real robots and synthetic large-scale workloads. The results show that ROSA improves factory productivity by up to 12.06× over conventional dedicated serving systems.  \n1 Introduction  \nEmbodied AI is emerging as a central frontier in the next phase of AI, promising physical agents that can perceive complex environments, reason over long-horizon tasks, and execute precise actions in the real world. Central to this shift are robotics foundation models (RFMs), including Vision-LanguageAction (VLA) models [6, 8, 22, 51, 52], World Action Models (WAMs) [1, 47, 50], and the surrounding reasoning and control models that support robotic execution. By integrating semantic language understanding with sensor inputs in the action-generation loop, these models have demonstrated  \nFigure 1: Robots working on various tasks in a factory.  \nunprecedented generalization capabilities across diverse tasks, from manipulation to navigation [5, 6, 16, 18, 35, 40, 44] .  \nA promising use case for RFMs is the deployment of general-purpose robots, such as humanoids and manipulators, in highly-automated robot factories. For example, Figure AI has explored humanoid deployments in BMW manufacturing facilities for automotive assembly-line tasks [14], Tesla is pursuing a similar vision with Optimus [41], and Amazon continues to scale warehouse robot fleets for inventory movement, sortation, and package handling [2, 33] .  \nRFM serving systems are central to this vision, since such deployments rely on continuous streams of model inference to generate robot actions and monitor task execution. Although recent work has begun to optimize RFM inference from both algorithmic and systems perspectives [6, 11, 30, 36, 47], we argue that the RFM serving problem for robot factories is neither well defined nor well solved, due in part to two pervasive misconceptions:  \nMisconception 1: robotics foundation model serving is strictly an edge computing problem. The prevailing deployment paradigm couples each robot with an onboard System-on-Chip (SoC) or a dedicated GPU server. However, one SoC per robot is not only expensive but also provides limited performance because robot platforms often have tight power and thermal budgets [17] . Offloading inference to a dedicated GPU server leads to better performance [19], but this one-to-one serving paradigm still under-utilizes GPUs: a GPU  \nserving only one robot cannot exploit inter-robot ","cbCaibgmR1MuUGBI","https://ap.wps.com/l/cbCaibgmR1MuUGBI","pdf",5092334,1,16,"English","en",105,"# Abstract\n# Introduction\n## Embodied AI and robotics foundation models\n## Robot-factory use cases\n## Why existing serving systems fall short\n### Misconception 1: edge-only serving\n### Misconception 2: latency-only objectives","[{\"question\":\"What limitations exist in current robotics foundation model serving systems for robot factories?\",\"answer\":\"They largely assume single-robot, single-model deployments and optimize mainly for single-action latency, which under-utilizes GPUs and ignores broader requirements like planning, safety checking, and monitoring models.\"},{\"question\":\"How does ROSA improve performance for fleets of robots?\",\"answer\":\"ROSA uses shared GPU-pool serving so multiple robots access centralized server-class GPUs over the network, improving inference performance, battery duration, and GPU utilization.\"},{\"question\":\"What scheduling objective does ROSA optimize compared with conventional systems?\",\"answer\":\"ROSA uses factory-objective-driven scheduling to maximize SLO-qualified factory productivity rather than minimizing latency for an individual request.\"}]",1784181641,40,{"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":85,"head_meta":87,"extra_data":89,"updated_unix":27},"rosa-robotics-foundation-model-serving-system-for-robot-factories","",{"@graph":35,"@context":84},[36,53,67],{"@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/rosa-robotics-foundation-model-serving-system-for-robot-factories/82581/",4,{"url":51,"name":13,"@type":54,"author":55,"headline":13,"publisher":57,"fileFormat":60,"inLanguage":23,"description":14,"dateModified":61,"datePublished":61,"encodingFormat":60,"isAccessibleForFree":62,"interactionStatistic":63},"DigitalDocument",{"name":9,"@type":56},"Person",{"url":40,"name":58,"@type":59},"DocShare","Organization","application/pdf","2026-07-16",true,{"@type":64,"interactionType":65,"userInteractionCount":4},"InteractionCounter",{"@type":66},"ViewAction",{"@type":68,"mainEntity":69},"FAQPage",[70,76,80],{"name":71,"@type":72,"acceptedAnswer":73},"What limitations exist in current robotics foundation model serving systems for robot factories?","Question",{"text":74,"@type":75},"They largely assume single-robot, single-model deployments and optimize mainly for single-action latency, which under-utilizes GPUs and ignores broader requirements like planning, safety checking, and monitoring models.","Answer",{"name":77,"@type":72,"acceptedAnswer":78},"How does ROSA improve performance for fleets of robots?",{"text":79,"@type":75},"ROSA uses shared GPU-pool serving so multiple robots access centralized server-class GPUs over the network, improving inference performance, battery duration, and GPU utilization.",{"name":81,"@type":72,"acceptedAnswer":82},"What scheduling objective does ROSA optimize compared with conventional systems?",{"text":83,"@type":75},"ROSA uses factory-objective-driven scheduling to maximize SLO-qualified factory productivity rather than minimizing latency for an individual request.","https://schema.org",{"og:url":51,"og:type":86,"og:title":13,"og:site_name":58,"og:description":14},"article",{"robots":88,"canonical":51},"index,follow",{"doc_id":7,"site_id":24},{"code":4,"msg":5,"data":91},[92,96,100,104,109,114,118,121,126,129,133],{"id":20,"doc_module":4,"doc_module_name":45,"category_name":93,"show_sort_weight":94,"slug":95},"Story & Novel",90,"story-novel",{"id":46,"doc_module":4,"doc_module_name":45,"category_name":97,"show_sort_weight":98,"slug":99},"Literature",80,"literature",{"id":52,"doc_module":4,"doc_module_name":45,"category_name":101,"show_sort_weight":102,"slug":103},"Exam",70,"exam",{"id":105,"doc_module":4,"doc_module_name":45,"category_name":106,"show_sort_weight":107,"slug":108},5,"Comic",60,"comic",{"id":110,"doc_module":4,"doc_module_name":45,"category_name":111,"show_sort_weight":112,"slug":113},6,"Technology",50,"technology",{"id":115,"doc_module":4,"doc_module_name":45,"category_name":116,"show_sort_weight":28,"slug":117},7,"Healthcare","healthcare",{"id":11,"doc_module":4,"doc_module_name":45,"category_name":12,"show_sort_weight":119,"slug":120},30,"research-report",{"id":122,"doc_module":4,"doc_module_name":45,"category_name":123,"show_sort_weight":124,"slug":125},9,"Religion & Spirituality",20,"religion-spirituality",{"id":124,"doc_module":4,"doc_module_name":45,"category_name":127,"show_sort_weight":124,"slug":128},"World Cup","world-cup",{"id":130,"doc_module":4,"doc_module_name":45,"category_name":131,"show_sort_weight":130,"slug":132},10,"Lifestyle","lifestyle",{"id":134,"doc_module":4,"doc_module_name":45,"category_name":135,"show_sort_weight":105,"slug":136},19,"General","general"]