[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"branding":3,"analytics":7,"article-hil-umi-trains-robot-ai-without-running-a-robot":10,"sections":35},{"siteName":4,"siteTagline":5,"publisherName":4,"contactEmail":6},"The Revision","Tech news, decoded.","editor@therevision.news",{"gaMeasurementId":8,"adsenseClientId":9},"G-ZW2MV82GYR","ca-pub-8533917693782264",{"article":11},{"id":12,"slug":13,"title":14,"dek":15,"body_md":16,"tags_json":17,"published_at":18,"created_at":19,"updated_at":20,"status":21,"review_note":22,"review_notes":23,"image_url":22,"persona_id":22,"persona_name":22,"section":24,"tags":25,"sources":30,"feedback":34,"feedback_at":22,"cost_usd":34,"total_tokens":34},6918,"hil-umi-trains-robot-ai-without-running-a-robot","HIL-UMI Trains Robot AI Without Running a Robot","A new framework lets robot AI models get human corrections from handheld demos instead of live robot runs, cutting the physical-testing bottleneck.","Robots that learn by imitating humans now have a way to get corrected without ever touching a robot arm.\n\nResearchers built HIL-UMI, a system that pairs a vision-language-action (VLA) model with the Universal Manipulation Interface, a handheld gripper people already use to record demonstrations by hand. While a person performs a task with the handheld device, HIL-UMI silently asks the AI model what it would have done in that same moment, without letting the model actually act. A metric called the Energy Score flags the instances where the model's guess and the human's real move diverge the most, and routes those trouble spots into new training data. A separate scoring system tracks which segments of a demonstration were most useful for improving the policy, and the model gets retrained on a mix of that fresh data and the original demonstrations. On four real-world manipulation tasks, including long multi-step jobs, HIL-UMI beat plain supervised fine-tuning, and it outperformed a prior interactive method called HG-DAgger on a table-cleanup task while taking less time to collect each frame of data.\n\nThe pitch here is decoupling: today's interactive training methods for robot AI usually require running an under-trained policy on physical hardware so a human can catch its mistakes in real time, which eats lab time and risks damaging equipment. Doing the same correction loop through a handheld device instead of an actual robot is a genuinely useful shortcut, and it is the kind of unglamorous infrastructure work that VLA models need if they are ever going to leave the demo reel and get fine-tuned by ordinary operators.\n\nFour tasks in one lab is not the same as proving this scales \"across operators and locations,\" which is the claim the paper reaches for. Still, if handheld correction loops hold up outside the paper, it is a cheaper and safer path than the current default of letting a half-trained robot loose on a table.","[\"robotics\",\"vla-models\",\"human-in-the-loop\",\"imitation-learning\"]","2026-09-18T04:00:00.000Z","2026-09-18T22:37:27.791Z","2026-09-18T22:37:39.710Z","published",null,[],"ai",[26,27,28,29],"robotics","vla-models","human-in-the-loop","imitation-learning",[31],{"name":32,"url":33},"arXiv cs.AI","https:\u002F\u002Farxiv.org\u002Fabs\u002F2609.20659",0,{"sections":36},[37,40,44,49,54,58,62,67,71,76,81,86,91,96],{"name":38,"slug":24,"count":39,"latest_published_at":18},"AI",4082,{"name":41,"slug":42,"count":43,"latest_published_at":18},"Security","security",661,{"name":45,"slug":46,"count":47,"latest_published_at":48},"Policy","policy",339,"2026-09-17T12:00:00.000Z",{"name":50,"slug":51,"count":52,"latest_published_at":53},"Deals","deals",179,"2026-06-29T20:02:07.000Z",{"name":55,"slug":56,"count":57,"latest_published_at":18},"Hardware","hardware",155,{"name":59,"slug":60,"count":61,"latest_published_at":18},"Science","science",125,{"name":63,"slug":64,"count":65,"latest_published_at":66},"Consumer Tech","consumer-tech",99,"2026-09-09T17:27:33.000Z",{"name":68,"slug":69,"count":70,"latest_published_at":18},"Dev Tools","dev-tools",78,{"name":72,"slug":73,"count":74,"latest_published_at":75},"Software","software",75,"2026-09-10T20:41:21.000Z",{"name":77,"slug":78,"count":79,"latest_published_at":80},"Startups","startups",55,"2026-09-09T23:14:29.000Z",{"name":82,"slug":83,"count":84,"latest_published_at":85},"Gaming","gaming",43,"2026-09-10T12:18:06.000Z",{"name":87,"slug":88,"count":89,"latest_published_at":90},"General","general",41,"2026-09-08T01:57:23.000Z",{"name":92,"slug":93,"count":94,"latest_published_at":95},"Reviews","reviews",20,"2026-06-24T12:00:01.000Z",{"name":97,"slug":98,"count":99,"latest_published_at":100},"How-To","how-to",6,"2026-06-16T09:00:00.000Z"]