[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"doc-detail-84722-en":3,"doc-seo-84722-105":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},84722,549758252649,"Ivy","https://ap-avatar.wpscdn.com/avatar/8000253669c5317157?_k=1778319167496531819",8,"Research & Report","Agent-driven Long-tail Simulation for Autonomous Driving","Evaluating autonomous driving systems in closed-loop environments requires realistic, interactive simulation, but many existing simulators rely on log replay or rule-based agents, limiting behavioral diversity and long-tail scenario coverage. This paper introduces an agent-driven simulation framework that uses instruction-following large language models with a structured action interface to control surrounding road participants while keeping physical plausibility. It also proposes SemanticPlan, a benchmark that extends nuPlan scenes with diverse language-instructed interactive agents.","arXiv :2607 .0433 1v 1 [ cs .RO] 5 Jul 2026  \nAgent-driven Long-tail Simulation for Autonomous Driving  \nJunru Gu 1 Lijin Yang2 Jianing Huang2 Shu Liu2 Zhongzhan Huang2  \nHang Zhao 1†  \n1IIIS, Tsinghua University 2Bosch Research  \nAbstract: Evaluating autonomous driving systems in closed-loop settings requires realistic and interactive simulation, yet existing simulators largely rely on log replay or rule-based agents, limiting behavioral diversity and long-tail coverage. We propose an agent-driven simulation framework in which surrounding road participants are controlled by instruction-following large language models through a structured action interface, enabling intentional and reactive behaviors while preserving physical plausibility. Furthermore, we introduce SemanticPlan, a benchmark of closed-loop planning in long-tail and semantically rich scenarios that augment real nuPlan scenes with multiple interactive agents following diverse language instructions. Evaluation results show that state-of-the-art planners still struggle to consistently achieve safe and effective task completion, suggesting that these long-tail scenarios remain challenging. 1  \nKeywords: Autonomous Driving, Planning, Simulation  \n1 Introduction  \nTo accurately and comprehensively evaluate autonomous driving systems, a realistic and interactive simulation environment is essential. In particular, surrounding road participants should be able to respond dynamically to the actions of the autonomous system, enabling meaningful closed-loop interaction.  \nExisting simulation frameworks predominantly rely on log replay or simple rule-based models to simulate the behavior of surrounding road participants [1] . While these approaches are efficient, they fail to capture the complexity and diversity of real-world human behaviors and cannot adequately respond to the actions of the tested autonomous system. For example, in closed-loop simulation settings, nuPlan [2, 3] employs log replay for pedestrian motion and IDM [4] for surrounding vehicles, resulting in limited interactivity and realism.  \nIn addition, long-tail scenarios are inherently rare in real-world driving datasets. Although simulators such as CARLA [5] and interPlan [6] attempt to address this issue by introducing synthetic data or additional road participants, they still rely on rule-based motion logic. As a result, they do not reflect realistic human behaviors and cannot adequately cover diverse and complex motion trajectories.  \nTo address these issues, we propose an agent-driven simulation framework, where specific road participants are controlled by instruction-following agents instead of predefined rules. This approach enables the simulation of complex and rare scenarios, where each road participant behaves as a realistic human with individual intentions and reactions.  \n1Project page: [https://tsinghua-mars-lab.github.io/SemanticPlan](https://tsinghua-mars-lab.github.io/SemanticPlan)  \n†Corresponding to: [hangzhao@mail.tsinghua.edu.cn](hangzhao@mail.tsinghua.edu.cn)  \nFigure 1: Overview of long-tail agents and scenario types in SemanticPlan.  \nTo enable such simulation, we construct long-tail and semantically rich scenarios, named the SemanticPlan dataset, on top of real-world driving scenes from the nuPlan dataset. SemanticPlan is designed as a closed-loop planning task where each scenario evaluates whether an ego planner can complete the task safely while interacting with controlled agents. SemanticPlan comprises over 50 scenario types, each involving multiple interactive agents with diverse language instructions. By placing agents at appropriate positions in different traffic scenes, we finally produce over 230 scenarios. Figure 1 provides representative long-tail agents and scenario types in SemanticPlan.  \nIn summary, the main contributions of this paper are:  \n• We propose an agent-driven simulation framework for autonomous driving evaluation, where surrounding road participants are contro","cbCaiiMOomkOoDtY","https://ap.wps.com/l/cbCaiiMOomkOoDtY","pdf",1359770,1,18,"English","en",105,"# Introduction\n## Related Work\n## Traffic simulation","[{\"question\":\"Why do existing autonomous driving simulators struggle with long-tail closed-loop evaluation?\",\"answer\":\"They often rely on log replay or rule-based agents, which limits the diversity of behaviors and reduces how well surrounding participants can react to the ego system in interactive settings.\"},{\"question\":\"How does the proposed agent-driven simulation framework generate surrounding-agent behavior?\",\"answer\":\"Surrounding road participants are controlled by instruction-following large language models via a structured action interface, enabling intentional and reactive behaviors while preserving physical plausibility.\"},{\"question\":\"What is SemanticPlan and what does it benchmark?\",\"answer\":\"SemanticPlan is a closed-loop planning benchmark built on long-tail and semantically rich scenarios, extending nuPlan with multiple interactive agents driven by diverse language instructions to test whether an ego planner completes tasks safely.\"}]",1784197849,45,{"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},"agent-driven-long-tail-simulation-for-autonomous-driving","",{"@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/agent-driven-long-tail-simulation-for-autonomous-driving/84722/",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},"Why do existing autonomous driving simulators struggle with long-tail closed-loop evaluation?","Question",{"text":75,"@type":76},"They often rely on log replay or rule-based agents, which limits the diversity of behaviors and reduces how well surrounding participants can react to the ego system in interactive settings.","Answer",{"name":78,"@type":73,"acceptedAnswer":79},"How does the proposed agent-driven simulation framework generate surrounding-agent behavior?",{"text":80,"@type":76},"Surrounding road participants are controlled by instruction-following large language models via a structured action interface, enabling intentional and reactive behaviors while preserving physical plausibility.",{"name":82,"@type":73,"acceptedAnswer":83},"What is SemanticPlan and what does it benchmark?",{"text":84,"@type":76},"SemanticPlan is a closed-loop planning benchmark built on long-tail and semantically rich scenarios, extending nuPlan with multiple interactive agents driven by diverse language instructions to test whether an ego planner completes tasks safely.","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},[93,97,101,105,110,115,120,123,128,131,135],{"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":124,"doc_module":4,"doc_module_name":45,"category_name":125,"show_sort_weight":126,"slug":127},9,"Religion & Spirituality",20,"religion-spirituality",{"id":126,"doc_module":4,"doc_module_name":45,"category_name":129,"show_sort_weight":126,"slug":130},"World Cup","world-cup",{"id":132,"doc_module":4,"doc_module_name":45,"category_name":133,"show_sort_weight":132,"slug":134},10,"Lifestyle","lifestyle",{"id":136,"doc_module":4,"doc_module_name":45,"category_name":137,"show_sort_weight":106,"slug":138},19,"General","general"]