[{"data":1,"prerenderedAt":219},["ShallowReactive",2],{"news-\u002Fnews\u002Fagentic-data-flywheel":3,"news-en-\u002Fnews\u002Fagentic-data-flywheel":121},{"id":4,"title":5,"author":6,"body":7,"date":102,"description":103,"draft":104,"extension":105,"image":106,"links":107,"meta":108,"name":109,"navigation":110,"path":111,"seo":112,"stem":113,"tags":114,"__hash__":120},"news\u002Fnews\u002Fagentic-data-flywheel.md","Trainingsdaten ohne Fernsteuerung: unser Agentic Data Flywheel","Symbiosika Team",{"type":8,"value":9,"toc":94},"minimark",[10,14,18,21,26,34,62,66,74,77,80,84,87,91],[11,12,5],"h1",{"id":13},"trainingsdaten-ohne-fernsteuerung-unser-agentic-data-flywheel",[15,16,17],"p",{},"Wer einem Roboter eine neue Aufgabe beibringen will, braucht Vorführungen, oft Hunderte. Üblich ist, dass Menschen den Roboter dafür per Fernsteuerung Schritt für Schritt durch die Aufgabe führen. Das ist langsam, teuer und lässt sich kaum skalieren.",[15,19,20],{},"Wir haben uns gefragt: Was wäre, wenn der Roboter sich seine Lehrstunden selbst schreibt?",[22,23,25],"h2",{"id":24},"eine-pipeline-die-sich-selbst-aufgaben-stellt","Eine Pipeline, die sich selbst Aufgaben stellt",[15,27,28,29,33],{},"Mit dem ",[30,31,32],"strong",{},"Agentic Data Flywheel"," haben wir eine eigene Pipeline gebaut, in der ein Team aus KI-Agenten die komplette Arbeit übernimmt, die sonst Menschen erledigen:",[35,36,37,44,50,56],"ol",{},[38,39,40,43],"li",{},[30,41,42],{},"Der Ideengeber"," überlegt sich neue Aufgaben: einen Becher vom Tisch nehmen, ein Handtuch ins Regal legen, eine Schublade öffnen.",[38,45,46,49],{},[30,47,48],{},"Der Bühnenbauer"," erschafft für jede Aufgabe Hunderte virtuelle Welten. Mal steht der Becher links, mal rechts, mal ist es hell, mal dämmrig, mal stehen andere Gegenstände im Weg.",[38,51,52,55],{},[30,53,54],{},"Der Lehrer"," schreibt die Strategie, mit der der Roboter die Aufgabe löst, als Programm.",[38,57,58,61],{},[30,59,60],{},"Der Prüfer"," schaut sich jeden Versuch unbestechlich an. Nur was wirklich gelungen ist, wird zu Trainingsdaten.",[22,63,65],{"id":64},"das-besondere-die-pipeline-lernt-dazu","Das Besondere: Die Pipeline lernt dazu",[15,67,68,69,73],{},"Hier beginnt der eigentliche Zauber. Scheitert ein Versuch, wird der Fehler nicht einfach verworfen. Der Lehrer analysiert, ",[70,71,72],"em",{},"warum"," der Griff danebenging, schreibt seine eigene Strategie um und probiert es erneut.",[15,75,76],{},"Mit jeder Runde wird der Generator also nicht nur produktiver, sondern klüger. Aufgaben, an denen er gestern noch gescheitert ist, löst er heute zuverlässig, und aus jeder gelösten Aufgabe entstehen Daten für die nächste. Das ist das Schwungrad: Je länger es läuft, desto schneller dreht es sich.",[15,78,79],{},"Während unser Team abends nach Hause geht, arbeitet die Pipeline weiter. Am nächsten Morgen liegen neue, geprüfte Vorführungen bereit, ohne dass jemand einen Joystick angefasst hat.",[22,81,83],{"id":82},"unabhängig-von-der-hardware","Unabhängig von der Hardware",[15,85,86],{},"Entwickelt haben wir das Flywheel für einen mobilen Manipulator. Gerade übertragen wir es auf einen zweiarmigen Roboter, ohne die Pipeline neu zu bauen. Das zeigt: Das Flywheel ist nicht an einen bestimmten Roboter gebunden. Es lernt einfach mit, was der jeweilige Körper kann.",[22,88,90],{"id":89},"warum-das-für-unseren-roboter-zählt","Warum das für unseren Roboter zählt",[15,92,93],{},"Jede neue Aufgabe im Hotel oder in der Pflege braucht Trainingsdaten. Wenn sie automatisch entstehen, wird aus wochenlanger Handarbeit ein Prozess, der über Nacht läuft. Genau so wollen wir neue Fähigkeiten schnell und günstig auf die Straße bringen.",{"title":95,"searchDepth":96,"depth":96,"links":97},"",2,[98,99,100,101],{"id":24,"depth":96,"text":25},{"id":64,"depth":96,"text":65},{"id":82,"depth":96,"text":83},{"id":89,"depth":96,"text":90},"2026-08-25T00:00:00.000Z","Der größte Engpass beim Roboterlernen sind Trainingsdaten. Unsere Pipeline erzeugt sie selbst und wird dabei mit jeder Runde besser.",false,"md","\u002Fimages\u002Flanding\u002Fband-lager.webp",null,{},"agentic-data-flywheel",true,"\u002Fnews\u002Fagentic-data-flywheel",{"title":5,"description":103},"news\u002Fagentic-data-flywheel",[115,116,117,118,119],"Robotik","KI","Simulation","Trainingsdaten","KI-Agenten","C8MIXH0GLWDqW2jQSt-RceFouNDWuifbWtZcGqo5WP0",{"id":122,"title":123,"author":6,"body":124,"date":102,"description":208,"draft":104,"extension":105,"image":106,"links":107,"meta":209,"name":109,"navigation":110,"path":210,"seo":211,"stem":212,"tags":213,"__hash__":218},"newsEn\u002Fnews-en\u002Fagentic-data-flywheel.md","Training data without teleoperation: our Agentic Data Flywheel",{"type":8,"value":125,"toc":202},[126,129,132,135,139,145,171,175,182,185,188,192,195,199],[11,127,123],{"id":128},"training-data-without-teleoperation-our-agentic-data-flywheel",[15,130,131],{},"Teaching a robot a new task takes demonstrations, often hundreds of them. The usual way is for people to guide the robot through the task step by step by remote control. That is slow, expensive and hard to scale.",[15,133,134],{},"So we asked ourselves: what if the robot wrote its own lessons?",[22,136,138],{"id":137},"a-pipeline-that-sets-its-own-tasks","A pipeline that sets its own tasks",[15,140,141,142,144],{},"With the ",[30,143,32],{}," we have built our own pipeline in which a team of AI agents does all the work that people would otherwise do:",[35,146,147,153,159,165],{},[38,148,149,152],{},[30,150,151],{},"The inventor"," comes up with new tasks: take a cup off the table, put a towel on the shelf, open a drawer.",[38,154,155,158],{},[30,156,157],{},"The stage builder"," creates hundreds of virtual worlds for each task. Sometimes the cup is on the left, sometimes on the right; sometimes it is bright, sometimes dim; sometimes other objects are in the way.",[38,160,161,164],{},[30,162,163],{},"The teacher"," writes the strategy the robot uses to solve the task as a program.",[38,166,167,170],{},[30,168,169],{},"The examiner"," judges every attempt impartially. Only what truly succeeded becomes training data.",[22,172,174],{"id":173},"what-makes-it-special-the-pipeline-learns","What makes it special: the pipeline learns",[15,176,177,178,181],{},"This is where the magic starts. When an attempt fails, the failure is not simply thrown away. The teacher works out ",[70,179,180],{},"why"," the grasp missed, rewrites its own strategy and tries again.",[15,183,184],{},"With every round, the generator becomes not just more productive but smarter. Tasks it failed at yesterday, it solves reliably today, and every solved task produces data for the next one. That is the flywheel: the longer it runs, the faster it spins.",[15,186,187],{},"When our team goes home in the evening, the pipeline keeps working. The next morning, new verified demonstrations are waiting, without anyone touching a joystick.",[22,189,191],{"id":190},"independent-of-hardware","Independent of hardware",[15,193,194],{},"We developed the flywheel for a mobile manipulator. Right now we are transferring it to a two-armed robot without rebuilding the pipeline. That shows the flywheel is not tied to one robot. It simply learns what each body can do.",[22,196,198],{"id":197},"why-this-matters-for-our-robot","Why this matters for our robot",[15,200,201],{},"Every new task in a hotel or in care needs training data. If it is generated automatically, weeks of manual work turn into a process that runs overnight. That is how we want to bring new skills to market quickly and affordably.",{"title":95,"searchDepth":96,"depth":96,"links":203},[204,205,206,207],{"id":137,"depth":96,"text":138},{"id":173,"depth":96,"text":174},{"id":190,"depth":96,"text":191},{"id":197,"depth":96,"text":198},"The biggest bottleneck in robot learning is training data. Our pipeline generates it on its own and gets better with every round.",{},"\u002Fnews-en\u002Fagentic-data-flywheel",{"title":123,"description":208},"news-en\u002Fagentic-data-flywheel",[214,215,117,216,217],"Robotics","AI","Training data","AI agents","GQ7o6Eowxc_qT467M4po_ixIBVUOl9tptjoMlX9Zlo0",1792586944992]