KGeN 🟩 (@KGeN_IO): Teaching Robots How To Cook

Teaching Robots How To Cook Humanoid robots are getting closer to entering our homes. But before they can fold laundry, wash dishes, or navigate a cluttered kitchen, they need to learn from someone who already does. Not from a simulation, lab demonstration, or a staged studio recording, but from a real person, doing real things. The footage that’s captured from a first-person, eye-level perspective, exactly as a human experiences a task, is called egocentric data. A robot needs to learn dexterity from the same vantage point it will eventually operate from. The Problem With How It’s Done Today Most robot training data is collected the wrong way. Currently, data is collected in staged environments, controlled lighting, and scripted movements performed for the camera. The result is robots perform faultlessly in demos and fail the moment they encounter a kitchen that isn't perfectly organised, or when a task deviates even slightly from the environments they were trained on. Dexterity isn't learned from demonstrations. It's learned from continuous, first-person observation of how human hands actually move through real environments. How do they reach? How do they adjust mid-motion? How do they recover when something slips or shifts? That signal exists in exactly one place: real homes. Why Collecting This Dataset Is Harder Than It Sounds Egocentric motion capture at scale introduces constraints that no standard data collection method can meet. Environments couldn't be controlled. Kitchens differed in layout, lighting, counter height, and spatial constraints. Living rooms were cluttered differently. Bathrooms varied. Humanoid systems need to generalise across exactly this kind of environmental diversity. Participants wore head-mounted smartphones while performing everyday chores, but the system only worked if the recording felt completely natural. The moment people became conscious of the camera, their movements changed. Actions became slower, more deliberate, and less authentic. For humanoid robots, that distinction matters. The goal wasn’t to capture perfect demonstrations; it was to capture real human behaviour as it naturally happens inside real homes. Environments couldn't be controlled. Kitchens differed in layout, lighting, counter height, and spatial constraints. Living rooms were cluttered differently. Bathrooms varied. Humanoid systems need to generalise across exactly this kind of environmental diversity. Quality requirements were strict: Hands had to remain visible and engaged in intentional, continuous task execution throughout Tasks couldn't be staged or performed out of the routine context Faces, children, and any identifiable information had to be excluded at the point of capture, not after collection. Technical fidelity had to be maintained across thousands of distributed recording hours simultaneously It was tough for any studio to achieve this, and not many freelancer platforms had the geographic reach. No synthetic pipeline could replicate what real human motion in real homes actually looks like. The Humans Behind Humyn This is where we at KGeN showed the value of a real community network. They became the backbone of the entire project. Rather than deploying hardware centrally, which would have made scaling impossible, KGeN's model turned participants themselves into the infrastructure. Contributors across South Asia, Southeast Asia, and the Middle East sourced compatible head-straps and smartphones locally and were reimbursed after verification. This eliminated shipping bottlenecks. The contributors also weren't random participants. They were domestic blue-collar workers who performed these tasks every single day as part of their actual lives: Cooking, cleaning, washing clothes, folding garments and maintaining living spaces. And this wasn't something we just coordinated from a distance. We did it ourselves, too! @mducasble, @KGeNBrazil lead, strapped on the headset and recorded himself performing his own household tasks, showing exactly how the process works, what it looks like in practice, and what it feels like to be a contributor. 📷 1M+ Active in Network. 30+ Countries . 20+ languages natively represented. The Result? 20,000+ hours of egocentric video of ordinary people performing their everyday tasks was collected across multiple countries under three months. Every contributor was sourced, verified, and routed through our AI vertical (KAI). Humyn Labs built the data infrastructure. KGeN supplied the real community.