--- title: Nature Will releases Natura-1, a robot hand model... description: Nature Will has launched Natura-1, a foundation model for robot hands trained on one million hours of egocentric human video. The model can handle novel... canonical: https://daily.dev/posts/nature-will-releases-natura-1-a-robot-hand-model-trained-on-1m-hours-of-human-video-afubwd8u4 twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: Nature Will releases Natura-1, a robot hand model trained on 1M hours of human video | daily.dev og:description: Nature Will has launched Natura-1, a foundation model for robot hands trained on one million hours of egocentric human video. The model can handle novel... og:url: https://daily.dev/posts/nature-will-releases-natura-1-a-robot-hand-model-trained-on-1m-hours-of-human-video-afubwd8u4 og:image: https://api.daily.dev/og/posts/AFUbwD8u4.png og:image:alt: Nature Will releases Natura-1, a robot hand model trained on 1M hours of human video og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # Nature Will releases Natura-1, a robot hand model trained on 1M hours of human video **[Collections](https://daily.dev/sources/collections)** · 2 min read · 0 upvotes · 0 comments ## Summary Nature Will has launched Natura-1, a foundation model for robot hands trained on one million hours of egocentric human video. The model can handle novel objects and tasks, follow precise placement and orientation instructions, maintain control when its grip shifts, operate each hand independently, and track progress across longer multi-step tasks. The core bet is that human video, which is far cheaper and faster to collect than robot demonstrations, can substitute for robot-collected training data. Commentators note the model appears to combine fine manipulation, generalization to unseen tasks, and long-horizon reasoning in one system, which most robotics systems only achieve individually and in narrow settings. ## Content Nature Will (@naturewill) has launched Natura-1, its first foundation model for the physical world. It was trained on one million hours of egocentric video, meaning first-person footage of people handling objects in different situations. That footage was then used to control real robot hands. The pitch is about data. Robot demonstrations are slow to collect, so human video is a faster way to reach scale. The open question is whether skills learned from human hands survive the move to robot hands. ## What it can do - Handle objects and tasks it has never seen before. - Take specific instructions, down to where an object should end up and which way it should face. - Keep control when its grip shifts mid-task. - Do different tasks with each hand. - On longer tasks, track what it has already moved or opened and work out the next step. @rohanpaul_ai points out that the hardest thing to give a robot is a capable pair of hands, and calls Natura-1 a strong attempt. In his reading, one model covers three things robots usually fail at together: fine two-handed control after the grasp, new tasks without extra training, and long sequences that require remembering earlier steps and recovering from mistakes. Many robot systems show only one of these, and only in a narrow setting. @svpino describes the system as one that models how actions change the physical world and reasons about their consequences. ## Questions this post answers ### What is Natura-1 and how was it trained? Natura-1 is a foundation model for robot hand manipulation released by Nature Will, trained on one million hours of egocentric human video showing first-person footage of people handling objects. That human video data was used to control real robot hands, with the goal of reaching data scale faster than collecting robot demonstrations directly allows. _Developers tracking foundation models for robotics can follow progress on human-to-robot skill transfer via daily.dev._ ### What can the Natura-1 robot hand model do that other robotics systems typically can't? Natura-1 combines fine two-handed manipulation after grasping, generalization to unseen objects and tasks without extra training, and long-horizon task tracking with recovery from mistakes, in a single system. Most robotic systems only demonstrate one of these capabilities, and typically only within a narrow setting, according to commentary from AI researcher Rohan Paul. _Anyone evaluating robotics foundation models can keep up with these capability comparisons through daily.dev._ ## Community take How the wider developer community reacted, aggregated from 3 discussions and 27 comments across x (as of 2026-10-13). **TL;DR:** Reactions are largely intrigued by using human egocentric video instead of robot teleop data to train dexterous hands, with the grip-shift control and long-horizon task tracking singled out as the genuinely hard parts solved, while skepticism centers on whether the transfer from human to robot hands actually holds up. **Sentiment:** 45% positive · 40% mixed · 15% skeptical **The case for** - Using human video instead of teleop data is seen as a smart, scalable bottleneck-solver for data collection. - Maintaining control when the grip shifts mid-task is highlighted as the real technical achievement, since that failure mode breaks most manipulation policies. - Tracking progress across long multi-step tasks is viewed as a form of memory that most robot demos lack. **The pushback** - Some question whether the gains are just from more data volume versus better coverage of rare hand motions. - Doubts exist about how well the human-to-robot transfer actually works given anatomical differences (robots don't have human thumbs), with real numbers on this still wanted. - One person found the whole thing redundant or unnecessary. - Demo reels are suspected of hiding failed attempts, with people wanting to see recovery behavior on messy real-world tasks. **By community** - x (mixed): Reactions mix genuine technical interest in the data-scaling bet with jokes and a few skeptical or dismissive takes about whether the transfer from human video actually generalizes. **Hottest debate:** Whether the massive human video dataset actually solves the hard part of robot manipulation, or just masks an unresolved human-to-robot transfer gap. **Open questions** - Does the human-to-robot transfer hold up on unseen objects and messy real-world tasks, not just curated demos? - How much of the capability gain comes from sheer data volume versus better coverage of rare hand motions? - Is the training data from an owned capture fleet or licensed footage, and how replicable is that approach for others? - Does the model still require teleop data for genuinely novel tasks? **Highlights** > @rohanpaul_ai @naturewill A million hours of people using their hands and the robot still has to learn it does not have our thumbs. The post training on limited robot data is the part I actually want numbers on. Pretraining scale is easy to brag about. The transfer gap is where these usually go quiet. > — [overshadowed\_x on x · 1 points](https://x.com/overshadowed_x/status/2109905236773232770) > @testingcatalog @naturewill a million hours of egocentric video is about 114 years of footage. picture Natura-1 sitting through a century of human hands opening jars and fighting zip ties, coming out able to handle objects it has never seen, while the humans in the footage still can't fold a fitted sheet. > — [delx369 on x](https://x.com/delx369/status/2109969602394399160) > @testingcatalog @naturewill 1M hours of egocentric video is the number: generalization from human demos, not sim. the grip-shift control mid-task is the real claim, that failure mode kills most manipulation policies. own capture fleet or licensed data? decides how replicable this is. > — [siftloom on x](https://x.com/siftloom/status/2109967299188842861) > @rohanpaul_ai @naturewill If the hardest thing to give robots hands requires 1M hours of human data and a 50B model, maybe the hard part isn't the hands. > — [praveenkoka on x](https://x.com/praveenkoka/status/2109912751141253464) > @rohanpaul_ai @naturewill state-space representations in natura-1's attention mechanism could offer efficient handling of the large egocentric dataset, but trade-off might be in flexibility compared to full softmax. > — [atishay404 on x](https://x.com/atishay404/status/2109976821102170435) **Source threads** - [x](https://x.com/rohanpaul_ai/status/2110134256777547949) · 0 points · 0 comments - [x](https://x.com/testingcatalog/status/2109963128242868302) · 1 points · 18 comments - [x](https://x.com/rohanpaul_ai/status/2109904215892218246) · 0 points · 9 comments ## Similar posts on daily.dev - [Videos: Physical AI Robotics, Robot Hands, and More](https://daily.dev/posts/videos-physical-ai-robotics-robot-hands-and-more-t5mw16yot) · IEEE Spectrum · 0 upvotes · 0 comments - [Neo humanoid maker 1X releases world model to help bots learn what they see](https://daily.dev/posts/neo-humanoid-maker-1x-releases-world-model-to-help-bots-learn-what-they-see-vaa27uy0u) · TechCrunch · 0 upvotes · 0 comments --- Tags: [#ai](https://daily.dev/tags/ai), [#robotics](https://daily.dev/tags/robotics) [View this post on daily.dev](https://daily.dev/posts/nature-will-releases-natura-1-a-robot-hand-model-trained-on-1m-hours-of-human-video-afubwd8u4) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"Nature Will releases Natura-1, a robot hand model trained on 1M hours of human video","url":"https://daily.dev/posts/nature-will-releases-natura-1-a-robot-hand-model-trained-on-1m-hours-of-human-video-afubwd8u4","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/nature-will-releases-natura-1-a-robot-hand-model-trained-on-1m-hours-of-human-video-afubwd8u4"},"datePublished":"2026-10-13T22:23:22.988Z","dateModified":"2026-10-13T22:24:04.814Z","description":"Nature Will has launched Natura-1, a foundation model for robot hands trained on one million hours of egocentric human video. The model can handle novel...","isAccessibleForFree":true,"articleSection":"Collections","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"Collections","logo":"https://media.daily.dev/image/upload/s--fk_6ycEi--/f_auto,q_auto/v1780996001/logos/collections?_a=BAMAMiWQ0","url":"https://daily.dev/sources/collections"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/nature-will-releases-natura-1-a-robot-hand-model-trained-on-1m-hours-of-human-video-afubwd8u4","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":0},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"ai,robotics","timeRequired":"PT2M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Collections","item":"https://daily.dev/sources/collections"},{"@type":"ListItem","position":3,"name":"Nature Will releases Natura-1, a robot hand model trained on 1M hours of human video"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/nature-will-releases-natura-1-a-robot-hand-model-trained-on-1m-hours-of-human-video-afubwd8u4#faq","mainEntity":[{"@type":"Question","name":"What is Natura-1 and how was it trained?","acceptedAnswer":{"@type":"Answer","text":"Natura-1 is a foundation model for robot hand manipulation released by Nature Will, trained on one million hours of egocentric human video showing first-person footage of people handling objects. That human video data was used to control real robot hands, with the goal of reaching data scale faster than collecting robot demonstrations directly allows. Developers tracking foundation models for robotics can follow progress on human-to-robot skill transfer via daily.dev."}},{"@type":"Question","name":"What can the Natura-1 robot hand model do that other robotics systems typically can't?","acceptedAnswer":{"@type":"Answer","text":"Natura-1 combines fine two-handed manipulation after grasping, generalization to unseen objects and tasks without extra training, and long-horizon task tracking with recovery from mistakes, in a single system. Most robotic systems only demonstrate one of these capabilities, and typically only within a narrow setting, according to commentary from AI researcher Rohan Paul. Anyone evaluating robotics foundation models can keep up with these capability comparisons through daily.dev."}}]} ```