{"id":1005141,"date":"2024-02-07T07:16:41","date_gmt":"2024-02-07T15:16:41","guid":{"rendered":"https:\/\/www.microsoft.com\/en-us\/research\/?post_type=msr-research-item&#038;p=1005141"},"modified":"2024-05-21T17:22:48","modified_gmt":"2024-05-22T00:22:48","slug":"interactive-agent-foundation-model","status":"publish","type":"msr-research-item","link":"https:\/\/www.microsoft.com\/en-us\/research\/publication\/interactive-agent-foundation-model\/","title":{"rendered":"An Interactive Agent Foundation Model"},"content":{"rendered":"<p><span dir=\"ltr\" role=\"presentation\">The development of artificial intelligence systems <\/span><span dir=\"ltr\" role=\"presentation\">is transitioning from creating static, task-specific<\/span><br role=\"presentation\" \/><span dir=\"ltr\" role=\"presentation\">models to dynamic, agent-based systems capa<\/span><span dir=\"ltr\" role=\"presentation\">ble of performing well in a wide range of ap<\/span><span dir=\"ltr\" role=\"presentation\">plications.<\/span> <span dir=\"ltr\" role=\"presentation\">We propose an<\/span> <strong><span dir=\"ltr\" role=\"presentation\">Agent <\/span><span dir=\"ltr\" role=\"presentation\">Foundation Model <\/span><\/strong><span dir=\"ltr\" role=\"presentation\">that uses a novel multi-task <\/span><span dir=\"ltr\" role=\"presentation\">agent training paradigm for training AI agents <\/span><span dir=\"ltr\" role=\"presentation\">across a wide range of domains, datasets, and <\/span><span dir=\"ltr\" role=\"presentation\">tasks. Our training paradigm unifies diverse pre<\/span><span dir=\"ltr\" role=\"presentation\">training strategies, including visual masked auto-<\/span><span dir=\"ltr\" role=\"presentation\">encoders, language modeling, and next-action <\/span><span dir=\"ltr\" role=\"presentation\">prediction, enabling a versatile and adaptable AI <\/span><span dir=\"ltr\" role=\"presentation\">framework. We demonstrate the performance of <\/span><span dir=\"ltr\" role=\"presentation\">our framework across three separate domains\u2014 <\/span><span dir=\"ltr\" role=\"presentation\">Robotics, Gaming AI, and Healthcare. Our model <\/span><span dir=\"ltr\" role=\"presentation\">demonstrates its ability to generate meaningful <\/span><span dir=\"ltr\" role=\"presentation\">and contextually relevant outputs in each area. <\/span><span dir=\"ltr\" role=\"presentation\">The strength of our approach lies in its general<\/span><span dir=\"ltr\" role=\"presentation\">ity, leveraging a variety of data sources such as <\/span><span dir=\"ltr\" role=\"presentation\">robotics sequences, gameplay data, large-scale <\/span><span dir=\"ltr\" role=\"presentation\">video datasets, and textual information for effec<\/span><span dir=\"ltr\" role=\"presentation\">tive multimodal and multi-task learning. Our ap<\/span><span dir=\"ltr\" role=\"presentation\">proach provides a promising avenue for develop<\/span><span dir=\"ltr\" role=\"presentation\">ing generalist, action-taking, multimodal systems.<\/span><\/p>\n","protected":false},"excerpt":{"rendered":"<p>The development of artificial intelligence systems is transitioning from creating static, task-specificmodels to dynamic, agent-based systems capable of performing well in a wide range of applications. We propose an Agent Foundation Model that uses a novel multi-task agent training paradigm for training AI agents across a wide range of domains, datasets, and tasks. Our training [&hellip;]<\/p>\n","protected":false},"featured_media":0,"template":"","meta":{"msr-url-field":"","msr-podcast-episode":"","msrModifiedDate":"","msrModifiedDateEnabled":false,"ep_exclude_from_search":false,"footnotes":""},"msr-content-type":[3],"msr-research-highlight":[],"research-area":[13556,13545,13554,13553],"msr-publication-type":[193716],"msr-product-type":[],"msr-focus-area":[],"msr-platform":[],"msr-download-source":[],"msr-locale":[268875],"msr-field-of-study":[268140,246658,268332,268266,247039,249835],"msr-conference":[],"msr-journal":[],"msr-impact-theme":[],"msr-pillar":[],"class_list":["post-1005141","msr-research-item","type-msr-research-item","status-publish","hentry","msr-research-area-artificial-intelligence","msr-research-area-human-language-technologies","msr-research-area-human-computer-interaction","msr-research-area-medical-health-genomics","msr-locale-en_us","msr-field-of-study-agent-ai","msr-field-of-study-deep-learning","msr-field-of-study-embodied-ai","msr-field-of-study-gaming","msr-field-of-study-health-care","msr-field-of-study-robotics"],"msr_publishername":"arXiv","msr_edition":"","msr_affiliation":"","msr_published_date":"2024-2-1","msr_host":"","msr_duration":"","msr_version":"","msr_speaker":"","msr_other_contributors":"","msr_booktitle":"","msr_pages_string":"","msr_chapter":"","msr_isbn":"","msr_journal":"","msr_volume":"","msr_number":"","msr_editors":"","msr_series":"","msr_issue":"","msr_organization":"","msr_how_published":"","msr_notes":"","msr_highlight_text":"","msr_release_tracker_id":"","msr_original_fields_of_study":"","msr_download_urls":"","msr_external_url":"","msr_secondary_video_url":"","msr_longbiography":"","msr_microsoftintellectualproperty":1,"msr_main_download":"","msr_publicationurl":"","msr_doi":"","msr_publication_uploader":[{"type":"file","viewUrl":"https:\/\/www.microsoft.com\/en-us\/research\/uploads\/prod\/2024\/02\/Agent-fondation-model.pdf","id":"1005228","title":"agent-fondation-model","label_id":"243109","label":0},{"type":"url","viewUrl":"false","id":"false","title":"https:\/\/arxiv.org\/pdf\/2402.05929.pdf","label_id":"243109","label":0}],"msr_related_uploader":[{"type":"file","viewUrl":"https:\/\/www.microsoft.com\/en-us\/research\/uploads\/prod\/2024\/02\/Agent-fondation-model.pdf","id":"1005228","title":"agent-fondation-model","label_id":"243118","label":0},{"type":"url","viewUrl":"false","id":"false","title":"https:\/\/arxiv.org\/pdf\/2402.05929.pdf","label_id":"243118","label":0}],"msr_attachments":[{"id":1005228,"url":"https:\/\/www.microsoft.com\/en-us\/research\/uploads\/prod\/2024\/02\/Agent-fondation-model.pdf"}],"msr-author-ordering":[{"type":"text","value":"Zane Durante","user_id":0,"rest_url":false},{"type":"text","value":"Bidipta Sarkar","user_id":0,"rest_url":false},{"type":"text","value":"Ran Gong","user_id":0,"rest_url":false},{"type":"text","value":"Rohan Taori","user_id":0,"rest_url":false},{"type":"guest","value":"yusuke-noda","user_id":969939,"rest_url":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/microsoft-research\/v1\/researchers?person=yusuke-noda"},{"type":"text","value":"Paul Tang","user_id":0,"rest_url":false},{"type":"text","value":"Ehsan Adeli","user_id":0,"rest_url":false},{"type":"text","value":"Shrinidhi Kowshika Lakshmikanth","user_id":0,"rest_url":false},{"type":"text","value":"Kevin Schulman","user_id":0,"rest_url":false},{"type":"text","value":"Arnold Milstein","user_id":0,"rest_url":false},{"type":"guest","value":"demetri-terzopoulos-2","user_id":981291,"rest_url":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/microsoft-research\/v1\/researchers?person=demetri-terzopoulos-2"},{"type":"user_nicename","value":"Ade Famoti","user_id":43005,"rest_url":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/microsoft-research\/v1\/researchers?person=Ade Famoti"},{"type":"text","value":"Noboru Kuno","user_id":0,"rest_url":false},{"type":"user_nicename","value":"Ashley Llorens","user_id":39964,"rest_url":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/microsoft-research\/v1\/researchers?person=Ashley Llorens"},{"type":"guest","value":"hoi-vo-3","user_id":981312,"rest_url":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/microsoft-research\/v1\/researchers?person=hoi-vo-3"},{"type":"user_nicename","value":"Katsushi Ikeuchi","user_id":32500,"rest_url":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/microsoft-research\/v1\/researchers?person=Katsushi Ikeuchi"},{"type":"guest","value":"fei-fei-li","user_id":969957,"rest_url":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/microsoft-research\/v1\/researchers?person=fei-fei-li"},{"type":"user_nicename","value":"Jianfeng Gao","user_id":32246,"rest_url":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/microsoft-research\/v1\/researchers?person=Jianfeng Gao"},{"type":"user_nicename","value":"Naoki Wake","user_id":39916,"rest_url":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/microsoft-research\/v1\/researchers?person=Naoki Wake"},{"type":"user_nicename","value":"Qiuyuan Huang","user_id":36356,"rest_url":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/microsoft-research\/v1\/researchers?person=Qiuyuan Huang"}],"msr_impact_theme":[],"msr_research_lab":[199565],"msr_event":[],"msr_group":[144931,668253],"msr_project":[788159],"publication":[],"video":[],"download":[],"msr_publication_type":"inproceedings","_links":{"self":[{"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-research-item\/1005141"}],"collection":[{"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-research-item"}],"about":[{"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/types\/msr-research-item"}],"version-history":[{"count":9,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-research-item\/1005141\/revisions"}],"predecessor-version":[{"id":1038630,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-research-item\/1005141\/revisions\/1038630"}],"wp:attachment":[{"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/media?parent=1005141"}],"wp:term":[{"taxonomy":"msr-content-type","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-content-type?post=1005141"},{"taxonomy":"msr-research-highlight","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-research-highlight?post=1005141"},{"taxonomy":"msr-research-area","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/research-area?post=1005141"},{"taxonomy":"msr-publication-type","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-publication-type?post=1005141"},{"taxonomy":"msr-product-type","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-product-type?post=1005141"},{"taxonomy":"msr-focus-area","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-focus-area?post=1005141"},{"taxonomy":"msr-platform","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-platform?post=1005141"},{"taxonomy":"msr-download-source","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-download-source?post=1005141"},{"taxonomy":"msr-locale","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-locale?post=1005141"},{"taxonomy":"msr-field-of-study","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-field-of-study?post=1005141"},{"taxonomy":"msr-conference","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-conference?post=1005141"},{"taxonomy":"msr-journal","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-journal?post=1005141"},{"taxonomy":"msr-impact-theme","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-impact-theme?post=1005141"},{"taxonomy":"msr-pillar","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-pillar?post=1005141"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}