Physics filtering favors the generalization of robot learning

Physics filtering favors the generalization of robot learning | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Article Physics filtering favors the generalization of robot learning Jindou Jia, Shixuan Han, Meng Wang, Gen Li, Zihan Yang, Sicheng Zhou, and 5 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-9460171/v1 This work is licensed under a CC BY 4.0 License Status: Under Review Version 1 posted 5 You are reading this latest preprint version Abstract Living organisms exhibit extraordinary adaptability to unseen environments through their intrinsic physical structures and lifelong feedback-driven learning. Endowing robots with comparable generalization is critical for reliable operation in the real world. While recent approaches attempt to improve generalization by scaling training data, such strategies remain impractical for robotics, where collecting real-world demonstrations at the scale of large language models is prohibitively costly and slow. Contrary to this reliance on massive datasets, we show that robots can generalize effectively even with limited training data by leveraging a feedback mechanism, namely PhyFilter, that corrects learning outputs with physics-filtered learning residuals. PhyFilter operates as a lightweight, model-agnostic module whose parameters can be automatically optimized through an auto-learning algorithm, eliminating manual tuning and enabling seamless integration with diverse robot policies. We validate PhyFilter across four representative robotic systems, demonstrating that it enables quadruped robots to generalize to unseen terrains, payload variations, and speed ranges; drones to flight under unseen wind disturbances; aerial manipulators to achieve centimeter-level in-air capture despite wind and mass uncertainties; and acceleration differentiators to remain robust with distribution shift. These results reveal a scalable and data-efficient pathway toward generalizable machine intelligence, showing that physics-filtered feedback can serve as a powerful alternative to massive data scaling. Physical sciences/Engineering Physical sciences/Mathematics and computing Physical sciences/Physics Full Text Additional Declarations No competing interests reported. Supplementary Files SupplementaryInformation.pdf SupplementaryMovies.zip Cite Share Download PDF Status: Under Review Version 1 posted Reviewers agreed at journal 13 May, 2026 Reviewers invited by journal 28 Apr, 2026 Editor assigned by journal 25 Apr, 2026 Submission checks completed at journal 23 Apr, 2026 First submitted to journal 19 Apr, 2026 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-9460171","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":630910907,"identity":"87a8371f-e22b-4201-b6c2-8c79af64e8b6","order_by":0,"name":"Jindou Jia","email":"","orcid":"","institution":"Nanyang Technological University","correspondingAuthor":false,"prefix":"","firstName":"Jindou","middleName":"","lastName":"Jia","suffix":""},{"id":630910908,"identity":"9712faee-595f-4cef-9e94-6951e1f9614b","order_by":1,"name":"Shixuan Han","email":"","orcid":"","institution":"Beihang University","correspondingAuthor":false,"prefix":"","firstName":"Shixuan","middleName":"","lastName":"Han","suffix":""},{"id":630910909,"identity":"9572b0e1-7b81-4911-ad20-30043daeb835","order_by":2,"name":"Meng Wang","email":"","orcid":"","institution":"Beihang University","correspondingAuthor":false,"prefix":"","firstName":"Meng","middleName":"","lastName":"Wang","suffix":""},{"id":630910910,"identity":"04beb85e-7ada-4914-a98e-b47a0940acd7","order_by":3,"name":"Gen Li","email":"","orcid":"","institution":"Nanyang Technological University","correspondingAuthor":false,"prefix":"","firstName":"Gen","middleName":"","lastName":"Li","suffix":""},{"id":630910911,"identity":"f5baaf77-386d-41fc-8ce5-b96f09db2700","order_by":4,"name":"Zihan Yang","email":"","orcid":"","institution":"Beihang University","correspondingAuthor":false,"prefix":"","firstName":"Zihan","middleName":"","lastName":"Yang","suffix":""},{"id":630910915,"identity":"f667d4f7-0319-4c1b-8dbc-849cf39c3605","order_by":5,"name":"Sicheng Zhou","email":"","orcid":"","institution":"Nanyang Technological University","correspondingAuthor":false,"prefix":"","firstName":"Sicheng","middleName":"","lastName":"Zhou","suffix":""},{"id":630910916,"identity":"f2af3841-54e7-4a84-bb2a-d57679a37c0a","order_by":6,"name":"Kexin Guo","email":"","orcid":"","institution":"Beihang University","correspondingAuthor":false,"prefix":"","firstName":"Kexin","middleName":"","lastName":"Guo","suffix":""},{"id":630910917,"identity":"2cc7e5d6-4283-4064-afdc-ab40abe407d6","order_by":7,"name":"Jianfei Yang","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAABCklEQVRIiWNgGAWjYBACxmYgUcEgAeF9gFAGhLWcgWphnAFRjV8LGJyB0sw8xGhhbmd+9uBAhQWDwfGzh1/b7viT2MDevE2C4U8tHoexmRscOCPBYHAmL80694xBYgPPsTIJxrbj+PxiJv2xDajlQI6ZcW4bUItEjpkEY8MxPFrYv0kc/AfUcv6NmbElSIv8GzOgw/Bp4TGTONgA1HIjx/gxI9gWoAgDWw0+LWUSB45J8EjeeGPG2NtmbNzGk1Zskdh2AKcWw/7j2yQO1NTJ8Z3PMf7ws01Otp/98MYbH/7U4dbSAKF5FA4wsIEjlA1EJDAcxqlFHs5oYGD+gCSB25ZRMApGwSgYcQAAFQdS1GZp4ogAAAAASUVORK5CYII=","orcid":"","institution":"Nanyang Technological University","correspondingAuthor":true,"prefix":"","firstName":"Jianfei","middleName":"","lastName":"Yang","suffix":""},{"id":630910918,"identity":"a035eab7-a325-4f8e-ad0b-037a3b3ac394","order_by":8,"name":"Xiang Yu","email":"","orcid":"","institution":"Beihang University","correspondingAuthor":false,"prefix":"","firstName":"Xiang","middleName":"","lastName":"Yu","suffix":""},{"id":630910919,"identity":"185b1dd8-83a7-4a26-b18a-e42b60beb670","order_by":9,"name":"Wei Wang","email":"","orcid":"","institution":"Beijing Aerospace Control Instrument Research Institute","correspondingAuthor":false,"prefix":"","firstName":"Wei","middleName":"","lastName":"Wang","suffix":""},{"id":630910920,"identity":"895dffa0-11ce-4c9d-a5a7-4b34403b7748","order_by":10,"name":"Lei Guo","email":"","orcid":"","institution":"Beihang University","correspondingAuthor":false,"prefix":"","firstName":"Lei","middleName":"","lastName":"Guo","suffix":""}],"badges":[],"createdAt":"2026-04-19 07:40:31","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-9460171/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-9460171/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":108809785,"identity":"5125ce10-1f58-410d-b04a-d3d551e23c17","added_by":"auto","created_at":"2026-05-08 15:55:28","extension":"pdf","order_by":1,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":5797521,"visible":true,"origin":"","legend":"","description":"","filename":"PhyFilterlatex.pdf","url":"https://assets-eu.researchsquare.com/files/rs-9460171/v1_covered_6c69a8a0-86c1-42a1-9d38-38ca2291fede.pdf"},{"id":108601490,"identity":"77aa4de1-7154-4cf6-a774-ceb6252631e6","added_by":"auto","created_at":"2026-05-06 11:33:30","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"supplement","size":2911900,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementaryInformation.pdf","url":"https://assets-eu.researchsquare.com/files/rs-9460171/v1/ea3fa2d7113dece334e045e6.pdf"},{"id":108805842,"identity":"7d2169f1-94a7-451d-91fa-b5f4bcd409cd","added_by":"auto","created_at":"2026-05-08 15:27:00","extension":"zip","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":139201943,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementaryMovies.zip","url":"https://assets-eu.researchsquare.com/files/rs-9460171/v1/ea1f6a6a11f0b1de1e08bd4a.zip"}],"financialInterests":"No competing interests reported.","formattedTitle":"Physics filtering favors the generalization of robot learning","fulltext":[],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":false,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":true,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":true,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"npj-robotics","isNatureJournal":false,"hasQc":false,"allowDirectSubmit":false,"externalIdentity":"npjrobot","sideBox":"Learn more about [npj Robotics](http://www.nature.com/npjrobot/)","snPcode":"44182","submissionUrl":"https://mts-npjrobot.nature.com/cgi-bin/main.plex","title":"npj Robotics","twitterHandle":"@NaturePortfolio","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"ejp","reportingPortfolio":"NPJ","inReviewEnabled":true,"inReviewRevisionsEnabled":true},"keywords":"","lastPublishedDoi":"10.21203/rs.3.rs-9460171/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-9460171/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eLiving organisms exhibit extraordinary adaptability to unseen environments through their intrinsic physical structures and lifelong feedback-driven learning. Endowing robots with comparable generalization is critical for reliable operation in the real world. While recent approaches attempt to improve generalization by scaling training data, such strategies remain impractical for robotics, where collecting real-world demonstrations at the scale of large language models is prohibitively costly and slow. Contrary to this reliance on massive datasets, we show that robots can generalize effectively even with limited training data by leveraging a feedback mechanism, namely PhyFilter, that corrects learning outputs with physics-filtered learning residuals. PhyFilter operates as a lightweight, model-agnostic module whose parameters can be automatically optimized through an auto-learning algorithm, eliminating manual tuning and enabling seamless integration with diverse robot policies. We validate PhyFilter across four representative robotic systems, demonstrating that it enables quadruped robots to generalize to unseen terrains, payload variations, and speed ranges; drones to flight under unseen wind disturbances; aerial manipulators to achieve centimeter-level in-air capture despite wind and mass uncertainties; and acceleration differentiators to remain robust with distribution shift. These results reveal a scalable and data-efficient pathway toward generalizable machine intelligence, showing that physics-filtered feedback can serve as a powerful alternative to massive data scaling.\u003c/p\u003e","manuscriptTitle":"Physics filtering favors the generalization of robot learning","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-05-06 11:33:16","doi":"10.21203/rs.3.rs-9460171/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"reviewerAgreed","content":"155710512325348453199538846855982974803","date":"2026-05-13T14:52:10+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2026-04-28T09:00:59+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2026-04-25T16:46:27+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2026-04-23T04:35:24+00:00","index":"","fulltext":""},{"type":"submitted","content":"npj Robotics","date":"2026-04-19T07:26:02+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"npj-robotics","isNatureJournal":false,"hasQc":false,"allowDirectSubmit":false,"externalIdentity":"npjrobot","sideBox":"Learn more about [npj Robotics](http://www.nature.com/npjrobot/)","snPcode":"44182","submissionUrl":"https://mts-npjrobot.nature.com/cgi-bin/main.plex","title":"npj Robotics","twitterHandle":"@NaturePortfolio","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"ejp","reportingPortfolio":"NPJ","inReviewEnabled":true,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"72a1af20-5bc4-4507-b164-acd503229c68","owner":[],"postedDate":"May 6th, 2026","published":true,"recentEditorialEvents":[{"type":"reviewerAgreed","content":"155710512325348453199538846855982974803","date":"2026-05-13T14:52:10+00:00","index":15,"fulltext":""}],"rejectedJournal":[],"revision":"","amendment":"","status":"under-review","subjectAreas":[{"id":67151729,"name":"Physical sciences/Engineering"},{"id":67151730,"name":"Physical sciences/Mathematics and computing"},{"id":67151731,"name":"Physical sciences/Physics"}],"tags":[],"updatedAt":"2026-05-06T11:33:17+00:00","versionOfRecord":[],"versionCreatedAt":"2026-05-06 11:33:16","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-9460171","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-9460171","identity":"rs-9460171","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}