[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"doc-detail-85535-en":3,"doc-seo-85535-105":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":20,"is_deleted":4,"is_public":21,"is_downloadable":21,"audit_status":21,"page_count":11,"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},85535,8796095360427,"Lucas Martin","https://ap-avatar.wpscdn.com/davatar_994ba38a5ba835b3df7d355c54d3ed8d",8,"Research & Report","Task Parameter Extrapolation via Learning Inverse Tasks from Forward Demonstrations","Generalizing skill policies to novel conditions remains a key challenge in robot learning. Imitation learning is data-efficient but often limited to the training region and can fail unpredictably on out-of-region inputs. Transfer learning may improve robustness but is frequently data-hungry and struggles with accurate zero-shot generalization. This work introduces a joint learning approach for task inversion learning by building a shared representation of forward and inverse tasks and using auxiliary forward demonstrations without direct inverse supervision. Experiments in simulation and real environments with complex manipulation outperform diffusion-based and multimodal VAE alternatives.","Task Parameter Extrapolation via Learning Inverse Tasks from  \nForward Demonstrations  \nSerdar Bahar1 , Fatih Dogangun 1 , Matteo Saveriano2 , Yukie Nagai3 , Emre Ugur1  \narXiv :2603 .05576v3 [ cs .RO] 13 Jul 2026  \nAbstract—Generalizing skill policies to novel conditions remains a key challenge in robot learning. Imitation learning methods, while data-efficient, are largely confined to the training region and consistently fail on input data outside it, leading to unpredictable policy failures. Alternatively, transfer learning approaches offer methods for trajectory generation robust to both changes in environment and tasks, but they remain datahungry and lack accuracy in zero-shot generalization. We address these challenges in the context of task inversion learning and propose a novel joint learning approach to achieve accurate and efficient knowledge transfer. Our method constructs a common representation of the forward and inverse tasks, and leverages auxiliary forward demonstrations from novel configurations to successfully execute the corresponding inverse tasks, without any direct supervision. We demonstrate the extrapolation capabilities of our framework through ablation studies and experiments in simulated and real-world environments that require complex manipulation skills with a diverse set of objects and tools, where we outperform diffusion-based and multimodal VAE alternatives.  \nIndex Terms—Learning from Experience, Learning from Demonstration, Joint Learning, Skill Generalization  \nI. INTRODUCTION  \nA fundamental challenge in robotics is developing methods for efficient, generalizable skill acquisition. An autonomous robot is expected to adapt to variations in its environment and changes in task or skill specifications. This requires skill policies that can generalize beyond training conditions, which remains a significant obstacle for current learning paradigms.  \nData-driven approaches in robotics have demonstrated impressive generalization capabilities, but often at the cost of extensive data collection and interaction with the environment [1] . Transfer learning frameworks aim to mitigate this by leveraging knowledge across tasks. A common technique, domain randomization, focuses on creating policies that are robust to variations in the environment, such as changes in lighting, friction, or object textures [2] . However, these approaches are mostly designed to handle environmental changes within a single task definition. While other methods, such as cross-domain transfer, exist, they often require additional training in the target domain and struggle with data efficiency and robust zero-shot performance [3] .  \nManuscript received: March, 6, 2026; Revised June, 3, 2026; Accepted June, 27, 2026 . This paper was recommended for publication by Editor T. Ogata upon evaluation of the Associate Editor and Reviewers’ comments. This work was supported by the EU through projects INVERSE (no. 101136067) and HARMONICA (no. 101294331), and in part by the World Premier International Research Center Initiative (WPI), MEXT, Japan.  \n1Department of Computer Engineering, Bogazici University, Istanbul, T¨urkiye. 2Department of Industrial Engineering, University of Trento, Trento, Italy. 3IRCN, the University of Tokyo, Tokyo, Japan. Corresponding author: [serdarbahar44@gmail.com](serdarbahar44@gmail.com)  \nCode is available at [https://github.com/serdarbahar/task](https://github.com/serdarbahar/task) par extrapolation  \nDigital Object Identifier (DOI): see top of this page.  \nImitation learning (IL) offers a more data-efficient alternative by leveraging expert demonstrations for the task of interest. Deep learning has given rise to powerful and flexible methods such as Conditional Neural Movement Primitives  \n[4] and Stable Movement Primitives [5], which learn complex, non-linear skills directly from demonstration data. More recently, advancements in deep generative modeling have enabled robotic policies to represent rich distr","cbCaidXekGKp6DBg","https://ap.wps.com/l/cbCaidXekGKp6DBg","pdf",9129572,3,1,"English","en",105,"# Introduction\n## Challenge of Skill Generalization in Robotics\n## Limitations of Imitation Learning and Existing Transfer Methods\n## Task Inversion and Joint Learning Framework","[{\"question\":\"Why do imitation learning policies often fail outside the training region?\",\"answer\":\"Imitation learning typically learns behaviors concentrated in the training configurations; when inputs fall outside that region, the learned policy can produce unpredictable trajectories that compromise downstream task success.\"},{\"question\":\"What problem does transfer learning address, and what is its drawback here?\",\"answer\":\"Transfer learning aims to generate robust trajectories under changes in environment and tasks, but it is often data-hungry and lacks accuracy for zero-shot generalization.\"},{\"question\":\"How does the proposed method learn inverse tasks without direct inverse supervision?\",\"answer\":\"The approach constructs a common representation for forward and inverse tasks and leverages auxiliary forward demonstrations from novel configurations to infer and execute corresponding inverse actions using sensorimotor similarity, without direct supervision for the inverse.\"}]",1784204271,20,{"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},"task-parameter-extrapolation-via-learning-inverse-tasks-from-forward-demonstrations","",{"@graph":35,"@context":84},[36,52,67],{"@type":37,"itemListElement":38},"BreadcrumbList",[39,43,47,49],{"item":40,"name":41,"@type":42,"position":21},"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":20},"https://docshare.wps.com/document/research-report/",{"item":50,"name":13,"@type":42,"position":51},"https://docshare.wps.com/document/task-parameter-extrapolation-via-learning-inverse-tasks-from-forward-demonstrations/85535/",4,{"url":50,"name":13,"@type":53,"author":54,"headline":13,"publisher":56,"fileFormat":59,"inLanguage":23,"description":14,"dateModified":60,"datePublished":61,"encodingFormat":59,"isAccessibleForFree":62,"interactionStatistic":63},"DigitalDocument",{"name":9,"@type":55},"Person",{"url":40,"name":57,"@type":58},"DocShare","Organization","application/pdf","2026-07-25","2026-07-16",true,{"@type":64,"interactionType":65,"userInteractionCount":20},"InteractionCounter",{"@type":66},"ViewAction",{"@type":68,"mainEntity":69},"FAQPage",[70,76,80],{"name":71,"@type":72,"acceptedAnswer":73},"Why do imitation learning policies often fail outside the training region?","Question",{"text":74,"@type":75},"Imitation learning typically learns behaviors concentrated in the training configurations; when inputs fall outside that region, the learned policy can produce unpredictable trajectories that compromise downstream task success.","Answer",{"name":77,"@type":72,"acceptedAnswer":78},"What problem does transfer learning address, and what is its drawback here?",{"text":79,"@type":75},"Transfer learning aims to generate robust trajectories under changes in environment and tasks, but it is often data-hungry and lacks accuracy for zero-shot generalization.",{"name":81,"@type":72,"acceptedAnswer":82},"How does the proposed method learn inverse tasks without direct inverse supervision?",{"text":83,"@type":75},"The approach constructs a common representation for forward and inverse tasks and leverages auxiliary forward demonstrations from novel configurations to infer and execute corresponding inverse actions using sensorimotor similarity, without direct supervision for the inverse.","https://schema.org",{"og:url":50,"og:type":86,"og:title":13,"og:site_name":57,"og:description":14},"article",{"robots":88,"canonical":50},"index,follow",{"doc_id":7,"site_id":24},{"code":4,"msg":5,"data":91},[92,96,100,104,109,114,119,122,126,129,133],{"id":21,"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":51,"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":117,"slug":118},7,"Healthcare",40,"healthcare",{"id":11,"doc_module":4,"doc_module_name":45,"category_name":12,"show_sort_weight":120,"slug":121},30,"research-report",{"id":123,"doc_module":4,"doc_module_name":45,"category_name":124,"show_sort_weight":28,"slug":125},9,"Religion & Spirituality","religion-spirituality",{"id":28,"doc_module":4,"doc_module_name":45,"category_name":127,"show_sort_weight":28,"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"]