[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"doc-detail-84762-en":3,"doc-seo-84762-105":28,"detail-sidebar-cat-0-en-105":89},{"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":11,"language":21,"language_code":22,"site_id":23,"html_lang":22,"table_of_contents":24,"faqs":25,"seo_title":13,"seo_description":14,"update_tm":26,"read_time":27},84762,4398048950312,"Violet","https://ap-avatar.wpscdn.com/avatar/400002538284de19e3c?_k=1778320343897328908",8,"Research & Report","Spatial Attention: Adapting Execution Horizons for Diffusion Policies via Observation Sensitivity","Sampling action chunks with generative diffusion models is widely used in robotic learning from demonstration, but fixed execution horizons create a persistent trade-off between responsiveness to disturbances and computational cost. This paper adaptively adjusts the execution horizon to balance both objectives. Spatial Attention is introduced as the expected squared norm of the gradient of the action log-likelihood with respect to the observation, capturing policy sensitivity to observation changes. With a fixed chunk-sampling budget, the horizon that minimizes cumulative likelihood drop decreases as Spatial Attention increases. A forecasting module assigns shorter horizons to high-attention phases and longer horizons to low-attention phases. Experiments in simulation and on a real aerial robot show higher success rates than fixed-horizon baselines while preserving the average horizon.","Spatial Attention: Adapting Execution Horizons for Diffusion Policies via Observation Sensitivity  \nChe-Sang Park, Junsu Ha, Jianlong Fu, and Frank C. Park  \narXiv :2607 .04739v 1 [ cs .RO] 6 Jul 2026  \nAbstract—Sampling action chunks via generative models has become a widely adopted methodology for robotic learning from demonstration. However, existing methods often struggle to balance responsiveness and computational cost because they execute each action chunk for a fixed execution horizon. In this paper, we adaptively adjust the execution horizon of sampled action chunks, balancing responsiveness and computational efficiency. We introduce Spatial Attention—defined as the expected squared norm of the gradient of the action log-likelihood with respect to the observation—which indicates the sensitivity of the policy’s action distribution to variations in the observation. We show that, under a fixed budget of chunk samplings, the execution horizon that minimizes the cumulative likelihood drop induced by disturbances decreases as Spatial Attention increases. By forecasting future Spatial Attention values alongside the action chunk, our framework dynamically assigns shorter execution horizons to phases with high Spatial Attention, and longer  \nhorizons to phases with low Spatial Attention. Experiments on standard and perturbed tasks, in both simulation and on areal robot, show that our method significantly improves success rates over fixed-horizon baselines while maintaining the average execution horizon.  \nIndex Terms—Machine Learning for Robot Control, Learning from Demonstration, Probabilistic Inference  \nI. INTRODUCTION  \nConsider a robotic system with state space X , observation space O, and action space A, governed by xk+1 = f (xk , ak ) + nk and ok = h (xk ) + wk , where xk ∈ X , ok ∈ O, and ak ∈ A denote the state, observation, and action at timestep k, and nk and wk are process and observation noise. The dynamics f and the observation function h vary with the embodiment and sensing modality. Since both fand h are unknown, learning from demonstration trains a policy π (a|o) that infers actions directly from observations, using a demonstration dataset D = { (oi , ai)}i=1 ,2··· . Recent robotic policies built on diffusion models [1], [2], [3] model π as a conditional generative model over action sequences. Given an observation ok , the policy samples a sequence of actions [ak , ak+1,..., ak+TH −1] ∼ π (·|ok), referred to as an action chunk [4] . The policy then executes only the first Ta (Ta ≤ TH ) actions of the chunk before sampling a new one, ina receding-horizon fashion similar to model predictive control; this Ta , referred to as the execution horizon, is fixed to a userspecified value prior to deployment.  \nExecuting an action chunk, however, leaves the policy  \nvulnerable to external disturbances and errors, since a new Under Review.  \nChe-Sang Park, Junsu Ha, and Frank C. Park are with the Robotics Laboratory, Seoul National University, Seoul, South Korea (e-mail: {cspark, [hajunsu](hajunsu}@robotics.snu.ac.kr)[}](hajunsu}@robotics.snu.ac.kr)[@robotics.snu.ac.kr](hajunsu}@robotics.snu.ac.kr); [fcp@snu.ac.kr](fcp@snu.ac.kr)).  \nJianlong Fu is with Microsoft Research Asia, Beijing, China (e-mail: [jianf@microsoft.com](jianf@microsoft.com)).  \nobservation is incorporated only after all Ta actions have been executed [5], [6] . Reducing the execution horizon allows more frequent observations and thus improves responsiveness, but each additional chunk sampling invokes the slow generative process, increasing computational cost and overall task completion time [2], [3], [7], [8], [9] . A fixed execution horizon therefore forces a single compromise between responsiveness and efficiency for the entire task, even though different phases of a task demand different levels of responsiveness.  \nThis motivates adapting the execution horizon during deployment, rather than fixing it in advance. Recently, several works in the domain o","cbCais9FUGPbT0oQ","https://ap.wps.com/l/cbCais9FUGPbT0oQ","pdf",1407748,1,"English","en",105,"# Abstract\n# Introduction","[{\"question\":\"What problem does the paper address in diffusion-based robot policies?\",\"answer\":\"Fixed execution horizons require each action chunk to be executed for a predetermined number of steps, creating a trade-off between responsiveness to disturbances and computational efficiency.\"},{\"question\":\"How is Spatial Attention defined and what does it represent?\",\"answer\":\"Spatial Attention is defined as the expected squared norm of the gradient of the action log-likelihood with respect to the observation, representing how sensitive the policy’s action distribution is to observation variations.\"},{\"question\":\"How does the method choose different execution horizons during deployment?\",\"answer\":\"A sequence-to-sequence estimator forecasts future Spatial Attention values and the policy dynamically assigns shorter horizons when accumulated predicted attention exceeds a threshold, and longer horizons when attention is low.\"}]",1784198092,20,{"code":4,"msg":29,"data":30},"ok",{"site_id":23,"language":22,"slug":31,"title":13,"keywords":32,"description":14,"schema_data":33,"social_meta":84,"head_meta":86,"extra_data":88,"updated_unix":26},"spatial-attention-adapting-execution-horizons-for-diffusion-policies-via-observation-sensitivity","",{"@graph":34,"@context":83},[35,52,66],{"@type":36,"itemListElement":37},"BreadcrumbList",[38,42,46,49],{"item":39,"name":40,"@type":41,"position":20},"https://docshare.wps.com","Home","ListItem",{"item":43,"name":44,"@type":41,"position":45},"https://docshare.wps.com/document/","Document",2,{"item":47,"name":12,"@type":41,"position":48},"https://docshare.wps.com/document/research-report/",3,{"item":50,"name":13,"@type":41,"position":51},"https://docshare.wps.com/document/spatial-attention-adapting-execution-horizons-for-diffusion-policies-via-observation-sensitivity/84762/",4,{"url":50,"name":13,"@type":53,"author":54,"headline":13,"publisher":56,"fileFormat":59,"inLanguage":22,"description":14,"dateModified":60,"datePublished":60,"encodingFormat":59,"isAccessibleForFree":61,"interactionStatistic":62},"DigitalDocument",{"name":9,"@type":55},"Person",{"url":39,"name":57,"@type":58},"DocShare","Organization","application/pdf","2026-07-16",true,{"@type":63,"interactionType":64,"userInteractionCount":4},"InteractionCounter",{"@type":65},"ViewAction",{"@type":67,"mainEntity":68},"FAQPage",[69,75,79],{"name":70,"@type":71,"acceptedAnswer":72},"What problem does the paper address in diffusion-based robot policies?","Question",{"text":73,"@type":74},"Fixed execution horizons require each action chunk to be executed for a predetermined number of steps, creating a trade-off between responsiveness to disturbances and computational efficiency.","Answer",{"name":76,"@type":71,"acceptedAnswer":77},"How is Spatial Attention defined and what does it represent?",{"text":78,"@type":74},"Spatial Attention is defined as the expected squared norm of the gradient of the action log-likelihood with respect to the observation, representing how sensitive the policy’s action distribution is to observation variations.",{"name":80,"@type":71,"acceptedAnswer":81},"How does the method choose different execution horizons during deployment?",{"text":82,"@type":74},"A sequence-to-sequence estimator forecasts future Spatial Attention values and the policy dynamically assigns shorter horizons when accumulated predicted attention exceeds a threshold, and longer horizons when attention is low.","https://schema.org",{"og:url":50,"og:type":85,"og:title":13,"og:site_name":57,"og:description":14},"article",{"robots":87,"canonical":50},"index,follow",{"doc_id":7,"site_id":23},{"code":4,"msg":5,"data":90},[91,95,99,103,108,113,118,121,125,128,132],{"id":20,"doc_module":4,"doc_module_name":44,"category_name":92,"show_sort_weight":93,"slug":94},"Story & Novel",90,"story-novel",{"id":45,"doc_module":4,"doc_module_name":44,"category_name":96,"show_sort_weight":97,"slug":98},"Literature",80,"literature",{"id":51,"doc_module":4,"doc_module_name":44,"category_name":100,"show_sort_weight":101,"slug":102},"Exam",70,"exam",{"id":104,"doc_module":4,"doc_module_name":44,"category_name":105,"show_sort_weight":106,"slug":107},5,"Comic",60,"comic",{"id":109,"doc_module":4,"doc_module_name":44,"category_name":110,"show_sort_weight":111,"slug":112},6,"Technology",50,"technology",{"id":114,"doc_module":4,"doc_module_name":44,"category_name":115,"show_sort_weight":116,"slug":117},7,"Healthcare",40,"healthcare",{"id":11,"doc_module":4,"doc_module_name":44,"category_name":12,"show_sort_weight":119,"slug":120},30,"research-report",{"id":122,"doc_module":4,"doc_module_name":44,"category_name":123,"show_sort_weight":27,"slug":124},9,"Religion & Spirituality","religion-spirituality",{"id":27,"doc_module":4,"doc_module_name":44,"category_name":126,"show_sort_weight":27,"slug":127},"World Cup","world-cup",{"id":129,"doc_module":4,"doc_module_name":44,"category_name":130,"show_sort_weight":129,"slug":131},10,"Lifestyle","lifestyle",{"id":133,"doc_module":4,"doc_module_name":44,"category_name":134,"show_sort_weight":104,"slug":135},19,"General","general"]