[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"doc-detail-85728-en":3,"doc-seo-85728-105":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},85728,4398048950312,"Violet","https://ap-avatar.wpscdn.com/avatar/400002538284de19e3c?_k=1778320343897328908",8,"Research & Report","TS-Mask VLA 2D Temporal Spatial Masking for Vision Language Action Model with Effective Bridging","Vision–language–action (VLA) models translate natural-language instructions and visual observations into executable robot actions, yet existing autoregressive token generation often collapses control into next-token prediction. This limits explicit spatiotemporal structure learning and disentanglement between vision-language representations and actions, especially in long-horizon, complex settings. TS-Mask VLA introduces a Discrete Diffusion Action Expert with a Bridge Attention conditioning bridge and a 2D temporal–spatial masking strategy for discrete action tokens, improving structural consistency. Experiments on LIBERO and CALVIN validate strong success rates, sequence length, and robust long-horizon performance, supported by analyses and ablations.","TS-Mask VLA: 2D Temporal–Spatial Masking for Vision-Language-Action Model with Effective Bridging  \nShengzhuo Yang 1∗, Ronghao Yu 1 , Chuanjie Lv 1 , Linpeng Peng 1 , Hang Yu2 , Jie Ren1 , Jiajun Lv 1 and Yong Liu 1†  \narXiv :2607 .09818v1 [ cs .RO] 10 Jul 2026  \nAbstract—Vision–language–action (VLA) models aim to understand natural-language instructions and visual observations, and to generate and execute corresponding actions as embodied agents. Recently, autoregressive token-based action generation has driven the development of many representative VLA models. However, this paradigm often reduces action generation to next-token prediction, thereby lacking explicit modeling of the spatiotemporal structure of action sequences and the disentanglement between visionlanguage representations and actions, which can limit performance in long-horizon and complex scenarios. In this paper, we propose TS-Mask VLA, a vision– language–action framework for robot manipulation. TS-Mask VLA is built upon two key designs: (1) We propose a Discrete Diffusion Action Expert equipped with a Bridge Attention conditioning bridge, which enables multi-layer conditioning from the VLM and facilitates more accurate and stable action generation; and(2) We propose a temporal–spatial 2D masking strategy for discrete action tokens that strengthens the model’s understanding of cross-time dependencies and inter-dimensional coupling, leading to more structurally consistent action sequences; and We conduct extensive experiments on simulation benchmarks and real-world tasks. On LIBERO, TS-Mask VLA achievesa 95.7% average success rate with only 0.5B parameters, outperforming significantly larger models. On CALVIN, it attains the best average sequence length of 4.19 and strong long-horizon performance. Comprehensive analyses and ablations further validate the effectiveness of our design.  \nI. INTRODUCTION  \nIn the past two years, rapid advances in multimodal large language models have made embodied robotic systems with general perception, understanding, and behavior a central research focus. Vision–Language–Action (VLA) models align visual perception, language understanding, and action generation within a unified framework, enabling instruction-driven robot manipulation. Most existing VLA build upon a pre-trained vision– language backbone with an action-generation head that maps observations and instructions to action sequences, following two dominant paradigms: (i) autoregressive  \n*This work was supported by “Zhejiang Key Laboratory of Advanced Intelligent Warehousing and Logistics Equipment” (Grant No. 2024E10007)  \n1 The authors are with the Institute of Cyber-Systems and Control, Zhejiang University, Hangzhou, China.  \n2 Hang Yu is with Tongji University, Shanghai, China.  \n∗ Equal contribution.  \n†Corresponding author (Email: [yongliu@iipc.zju.edu.cn](yongliu@iipc.zju.edu.cn)) .  \ntransformers that sequentially predict action tokens [1]–[3], and (ii) diffusion-based approaches that generate continuous trajectories via iterative denoising [4], [5] . Despite this progress, VLA still face challenges in longhorizon tasks and complex environments, particularly in balancing structural unification and efficiency.  \nCurrent VLA methods predominantly follow two paradigms, both exhibiting inherent limitations. The first unifies vision, language, and actions into a single token sequence and autoregressively generates actions [3], lacking explicit structural decoupling between representation learning and control policy modeling. The second adopts continuous diffusion to iteratively generate trajectories [6], which can lead to instability over long horizons and makes it difficult to explicitly model temporal–spatial correlations in robotic actions. Motivated by these observations, we argue that two aspects are crucial: (i) bridging vision-language representations into a dedicated action expert to establish clearer conditional decoupling; and (ii) discretizing actions to e","cbCaisDlpzc0qjQH","https://ap.wps.com/l/cbCaisDlpzc0qjQH","pdf",2406628,1,9,"English","en",105,"# Introduction\n# Method Overview\n## Discrete Diffusion Action Expert with Bridge Attention\n## Temporal–Spatial 2D Masking Strategy","[{\"question\":\"What problem does TS-Mask VLA target in existing VLA models?\",\"answer\":\"It addresses the tendency of token autoregression to reduce action generation to next-token prediction, which weakens explicit spatiotemporal structure modeling and representation-to-action disentanglement in long-horizon scenarios.\"},{\"question\":\"How does TS-Mask VLA incorporate vision-language information into the action generation process?\",\"answer\":\"It extracts visual features from multiple cameras, uses a lightweight VLM backbone, then injects intermediate VLM-layer features into a dedicated discrete diffusion action expert via a Bridge Attention conditioning bridge.\"},{\"question\":\"What is the role of the temporal–spatial 2D masking strategy?\",\"answer\":\"It discretizes actions into tokens organized by time steps and action dimensions, then applies structured masking along both temporal and action-dimension axes to strengthen cross-time dependencies and inter-dimensional coupling.\"}]",1784205854,23,{"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},"ts-mask-vla-2d-temporal-spatial-masking-for-vision-language-action-model-with-effective-bridging","",{"@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/ts-mask-vla-2d-temporal-spatial-masking-for-vision-language-action-model-with-effective-bridging/85728/",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 TS-Mask VLA target in existing VLA models?","Question",{"text":75,"@type":76},"It addresses the tendency of token autoregression to reduce action generation to next-token prediction, which weakens explicit spatiotemporal structure modeling and representation-to-action disentanglement in long-horizon scenarios.","Answer",{"name":78,"@type":73,"acceptedAnswer":79},"How does TS-Mask VLA incorporate vision-language information into the action generation process?",{"text":80,"@type":76},"It extracts visual features from multiple cameras, uses a lightweight VLM backbone, then injects intermediate VLM-layer features into a dedicated discrete diffusion action expert via a Bridge Attention conditioning bridge.",{"name":82,"@type":73,"acceptedAnswer":83},"What is the role of the temporal–spatial 2D masking strategy?",{"text":84,"@type":76},"It discretizes actions into tokens organized by time steps and action dimensions, then applies structured masking along both temporal and action-dimension axes to strengthen cross-time dependencies and inter-dimensional coupling.","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"]