[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"branding":3,"analytics":7,"article-a-ratio-trick-keeps-adversarial-rl-training-from-collapsing":10,"sections":36},{"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":24,"persona_id":22,"persona_name":22,"section":25,"tags":26,"sources":31,"feedback":35,"feedback_at":22,"cost_usd":35,"total_tokens":35},7110,"a-ratio-trick-keeps-adversarial-rl-training-from-collapsing","A Ratio Trick Keeps Adversarial RL Training From Collapsing","A new ratio-based method called CoDRA stabilizes adversarial RL training without hand-tuned penalties, tested on simulated pendulums.","Researchers have a new way to make reinforcement learning controllers tougher without wrecking their baseline performance.\n\nReinforcement learning controllers trained purely in simulation often stumble once they hit real hardware, because the real world throws disturbances the simulator never modeled. The standard fix, called robust RL, trains the controller alongside a simulated adversary that pushes and shoves it during training so it learns to shrug off surprises. The catch: an adversary that isn't reined in makes training unstable, and the resulting controller often performs worse than one that skipped robustness training altogether. A new method called CoDRA, short for cost-to-disturbance ratio adversarial training, fixes this by tracking the ratio between how much the controller struggles and how hard the adversary is pushing, then using that ratio, self-normalized batch by batch, to automatically throttle the adversary's incentive, no manual penalty term or tuning knob required.\n\nRobust RL has mostly been a tradeoff: make a controller resilient to disturbances and you usually pay for it with worse everyday performance, which is why plenty of robotics teams skip it. On two MuJoCo pendulum benchmarks, CoDRA's controller held the lowest cost across nearly every force and mass setting tested, including a force level it never saw during training, without the usual collapse in baseline performance. That's the actual selling point: not that it survives disturbances, but that survival doesn't cost you anything.\n\nIt's still lab-bench physics, a pair of pendulums, not a warehouse robot, so read the results as a promising proof of concept rather than a verdict.","[\"reinforcement-learning\",\"adversarial-training\",\"robotics\",\"ai-research\"]","2026-09-21T04:00:00.000Z","2026-09-21T07:02:40.505Z","2026-09-21T07:02:52.179Z","published",null,[],"https:\u002F\u002Fcdn.xyz.onl\u002Farticle-images\u002Fa-ratio-trick-keeps-adversarial-rl-training-from-collapsing.webp","ai",[27,28,29,30],"reinforcement-learning","adversarial-training","robotics","ai-research",[32],{"name":33,"url":34},"arXiv cs.AI","https:\u002F\u002Farxiv.org\u002Fabs\u002F2603.12110",0,{"sections":37},[38,42,46,51,56,61,66,71,76,81,86,91,96,101],{"name":39,"slug":25,"count":40,"latest_published_at":41},"AI",4175,"2026-09-21T10:30:00.000Z",{"name":43,"slug":44,"count":45,"latest_published_at":18},"Security","security",681,{"name":47,"slug":48,"count":49,"latest_published_at":50},"Policy","policy",352,"2026-09-21T10:18:06.000Z",{"name":52,"slug":53,"count":54,"latest_published_at":55},"Deals","deals",184,"2026-09-21T10:18:31.000Z",{"name":57,"slug":58,"count":59,"latest_published_at":60},"Hardware","hardware",157,"2026-09-21T11:04:12.000Z",{"name":62,"slug":63,"count":64,"latest_published_at":65},"Science","science",130,"2026-09-20T13:48:11.000Z",{"name":67,"slug":68,"count":69,"latest_published_at":70},"Consumer Tech","consumer-tech",99,"2026-09-09T17:27:33.000Z",{"name":72,"slug":73,"count":74,"latest_published_at":75},"Dev Tools","dev-tools",78,"2026-09-18T04:00:00.000Z",{"name":77,"slug":78,"count":79,"latest_published_at":80},"Software","software",75,"2026-09-10T20:41:21.000Z",{"name":82,"slug":83,"count":84,"latest_published_at":85},"Startups","startups",55,"2026-09-09T23:14:29.000Z",{"name":87,"slug":88,"count":89,"latest_published_at":90},"Gaming","gaming",43,"2026-09-10T12:18:06.000Z",{"name":92,"slug":93,"count":94,"latest_published_at":95},"General","general",42,"2026-09-18T22:35:10.000Z",{"name":97,"slug":98,"count":99,"latest_published_at":100},"Reviews","reviews",20,"2026-06-24T12:00:01.000Z",{"name":102,"slug":103,"count":104,"latest_published_at":105},"How-To","how-to",6,"2026-06-16T09:00:00.000Z"]