[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"detail-sidebar-cat-0-en-105":3,"doc-seo-327334-105":59,"doc-detail-327334-en":131},{"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":124,"head_meta":126,"extra_data":128,"updated_unix":130},105,"en","path-level-hindsight-instructions-for-semantic-exploration-in-vision-language-navigation","Path-level Hindsight Instructions for Semantic Exploration in Vision-Language Navigation","","On-policy exploration is essential for training robust Vision-Language Navigation (VLN) agents because it expands the state distribution they encounter. Yet exploratory trajectories often diverge from expert demonstrations, creating a semantic mismatch between the visual stream and the original language instruction. Φ-Nav introduces a unified on-policy framework that uses hindsight reasoning to realign instructions with the agent’s actual exploratory journey via a three-stage dual-supervision cycle and improves instruction-guided learning efficiency with limited expert data.",{"@graph":69,"@context":123},[70,84,106],{"@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/path-level-hindsight-instructions-for-semantic-exploration-in-vision-language-navigation/327334/",{"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/path-level-hindsight-instructions-for-semantic-exploration-in-vision-language-navigation/327334.png","ImageObject",300,407,{"name":92,"@type":93},"Emma Wilson","Person",{"url":74,"name":95,"@type":96},"DocShare","Organization","application/pdf","2026-10-24","2026-09-21",true,{"@type":102,"interactionType":103,"userInteractionCount":105},"InteractionCounter",{"@type":104},"ViewAction",14,{"@type":107,"mainEntity":108},"FAQPage",[109,115,119],{"name":110,"@type":111,"acceptedAnswer":112},"Why does on-policy exploration cause a semantic mismatch in Vision-Language Navigation?","Question",{"text":113,"@type":114},"On-policy exploration makes agents visit states not fully covered by expert trajectories. When the agent’s path deviates, the original instruction may no longer describe the executed visual stream, creating a semantic supervision gap.","Answer",{"name":116,"@type":111,"acceptedAnswer":117},"How does Φ-Nav generate path-level hindsight instructions?",{"text":118,"@type":114},"Φ-Nav first runs oracle-guided on-policy exploration to collect a trajectory with expert action feedback. Then a hindsight speaker synthesizes a path-level instruction grounded in the collected visual observations.",{"name":120,"@type":111,"acceptedAnswer":121},"What is the role of the second imitation pass in Φ-Nav?",{"text":122,"@type":114},"After synthesizing the trajectory–instruction pair, the agent performs a second imitation pass treating this pair as an additional expert demonstration, reinforcing semantic grounding along the executed path.","https://schema.org",{"og:url":83,"og:type":125,"og:title":65,"og:site_name":95,"og:description":67},"article",{"robots":127,"canonical":83},"index,follow",{"doc_id":129,"site_id":62},327334,1790387990,{"code":4,"msg":5,"data":132},{"doc_id":129,"user_id":133,"nickname":92,"user_avatar":134,"doc_module":4,"category_id":39,"category_name":40,"doc_title":65,"doc_description":67,"doc_content":135,"file_id":136,"file_url":137,"file_type":138,"file_size":139,"view_count":105,"is_deleted":4,"is_public":8,"is_downloadable":8,"audit_status":8,"page_count":140,"language":141,"language_code":63,"site_id":62,"html_lang":63,"table_of_contents":142,"faqs":143,"seo_title":144,"seo_description":67,"update_tm":145,"read_time":26,"preview_page_count":4},3848291630094,"https://eur-avatar.wpscdn.com/davatar_085a072bc5b1113ac321206ff7593b45","arXiv :2607 .0 1754v 1 [ cs .AI] 2 Jul 2026  \nPath-level Hindsight Instructions for Semantic Exploration in Vision-Language Navigation  \nSung June Kim 1 ,2⋆, Sangpil Kim 1†, and Honglak Lee2†  \n1 Korea University, Seoul, South Korea  \n2 University of Michigan, Ann Arbor, Michigan, USA  \nAbstract. On-policy exploration is a crucial component for training robust Vision-Language Navigation (VLN) agents, as it exposes the policy to a broader state distribution. However, such exploration inevitably leads to trajectories that deviate from expert demonstrations, resulting ina semantic mismatch between the executed visual stream and the original language instruction. In this work, we address this challenge by introducing Φ-Nav, a unified on-policy framework that leverages hindsight reasoning to align instructions with the agent’s actual exploratory journey.  \nSpecifically, Φ-Nav operates through a three-stage dual-supervision cycle: 1) the agent performs oracle-guided on-policy exploration, sampling a trajectory while learning from expert action feedback, 2) a hindsight speaker synthesizes a path-level hindsight instruction grounded in the collected visual observations, and 3) the agent conducts a second imitation pass, treating the synthesized trajectory–instruction pair as an additional expert demonstration. Through this process, Φ-Nav bridges the critical semantic supervision gap inherent in on-policy methods, transforming semantically unlabeled movement into dense training signals.  \nEvaluations on the R2R-CE and RxR-CE benchmarks show that Φ-Nav yields competitive performance while requiring only a fraction of the expert demonstrations used by current baselines. These results underscore the necessity of semantic exploration in VLN, positioning Φ-Nav as a effective solution for training embodied agents with limited data.  \nKeywords: Vision-language navigation · On-policy imitation learning · Hindsight experiential learning  \n1 Introduction  \nThe field of Vision-Language Navigation (VLN) is a cornerstone of embodied AI, requiring agents to ground natural language instructions in complex, previously unseen environments [3, 59] . To achieve robust generalization, modern VLN systems increasingly rely on on-policy exploration during training, exposing agents to diverse and dynamically evolving state distributions [2, 13, 32, 37, 62] . While architectural  advances have significantly improved perception and cross-modal  \n⋆ Work done at the University of Michigan as a visiting researcher.† Corresponing [authors {spk7@korea.ac.kr](authors {spk7@korea.ac.kr) ; [honglak@umich.edu}](honglak@umich.edu})  \n2 S. J. Kim et al.  \nFig. 1: Conceptual comparison of VLN training strategies. (a) Off-policy IL relies on static expert paths, failing to account for agent exploration. (b) On-policy IL exposes the agent to exploratory states but remains semantically limited to the original instruction, creating a supervision mismatch. (c) We utilize path-level hindsight relabeling to generate new instructions for exploratory trajectories, transforming deviations into meaningful training signals.  \nreasoning, the fundamental challenge of aligning exploratory behavior with coherent linguistic supervision remains unresolved.  \nTo encourage robustness, many VLN policies adopt on-policy strategies such as scheduled sampling [6] and DAgger [45], which mitigate exposure bias by allowing agents to act under their own policies while receiving expert action feedback 1. Although these methods broaden the explored state space, they remain semantically tethered to static, pre-collected instructions. For example, when an agent deviates from the expert’s intended route, the original instruction no longer faithfully describes the executed visual stream. As a result, a substantial portion of exploratory trajectories lacks semantically aligned linguistic supervision, thereby limiting the full utilization of on-policy trajectories.  \nTo bridge the semantic supervision gap in","cbCait15MjkTjmAW","https://ap.wps.com/l/cbCait15MjkTjmAW","pdf",2083296,24,"English","# Introduction\n## Path-level semantic supervision gap\n## Proposed Φ-Nav framework and cycle\n## Challenges of hindsight relabeling for open-ended trajectories","[{\"question\":\"Why does on-policy exploration cause a semantic mismatch in Vision-Language Navigation?\",\"answer\":\"On-policy exploration makes agents visit states not fully covered by expert trajectories. When the agent’s path deviates, the original instruction may no longer describe the executed visual stream, creating a semantic supervision gap.\"},{\"question\":\"How does Φ-Nav generate path-level hindsight instructions?\",\"answer\":\"Φ-Nav first runs oracle-guided on-policy exploration to collect a trajectory with expert action feedback. Then a hindsight speaker synthesizes a path-level instruction grounded in the collected visual observations.\"},{\"question\":\"What is the role of the second imitation pass in Φ-Nav?\",\"answer\":\"After synthesizing the trajectory–instruction pair, the agent performs a second imitation pass treating this pair as an additional expert demonstration, reinforcing semantic grounding along the executed path.\"}]","Path-level Hindsight Instructions for Semantic Exploration in Vision-Language Navigation | PDF",1789977002]