--- title: Dyna-2: General Purpose Physical Agents with Hour-Long... description: Dyna Robotics co-founder and Chief Scientist Jason Ma discusses Dyna-2, a world-action model pre-trained on over one million hours of human video, and the... canonical: https://daily.dev/posts/dyna-2-general-purpose-physical-agents-with-hour-long-autonomy-jason-ma-dyna-robotics-sittjfsm7 twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: Dyna-2: General Purpose Physical Agents with Hour-Long Autonomy — Jason Ma, Dyna Robotics | daily.dev og:description: Dyna Robotics co-founder and Chief Scientist Jason Ma discusses Dyna-2, a world-action model pre-trained on over one million hours of human video, and the... og:url: https://daily.dev/posts/dyna-2-general-purpose-physical-agents-with-hour-long-autonomy-jason-ma-dyna-robotics-sittjfsm7 og:image: https://api.daily.dev/og/posts/SittJfSM7.png og:image:alt: Dyna-2: General Purpose Physical Agents with Hour-Long Autonomy — Jason Ma, Dyna Robotics 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. # Dyna-2: General Purpose Physical Agents with Hour-Long Autonomy — Jason Ma, Dyna Robotics **[Latent Space](https://daily.dev/sources/latentspace)** · 73 min read · 0 upvotes · 0 comments ## Summary Dyna Robotics co-founder and Chief Scientist Jason Ma discusses Dyna-2, a world-action model pre-trained on over one million hours of human video, and the follow-up physical robot agent Dyna-2.1. The conversation covers how world-action models predict future video states to guide robot actions, the use of diffusion transformers and flow matching for smooth real-time control, and how robots can learn new tasks from as little as 45 minutes of demonstrations. Ma explains the data bottleneck in robotics compared to language models, the rationale for building custom hardware (wheeled semi-humanoid with torso movement) instead of relying on off-the-shelf humanoids, and current hardware limitations like overheating dexterous hands. Real-world deployments at Best Western and Din Tai Fung are discussed, where robots fold thousands of napkins daily at human parity and consistency, with Din Tai Fung now expanding the contract to include food-cutting tasks. The episode also touches on synthetic data generation, Ma's Eureka project background, on-device inference constraints, and a 'bitter lesson' framing for robotics scaling. ## Full article daily.dev links to this article rather than hosting it. Read it at the original source: ## Questions this post answers ### What is a world-action model in robotics and how does it differ from a vision-language-action model? A world-action model predicts future video or state of the physical world and uses that prediction to generate robot actions, similar to a chain-of-thought reasoning process in visual and physics space before acting. Dyna-2 is built on a diffusion transformer architecture trained on over one million hours of human video, combining future video prediction with robot action generation, making it more dexterous and better at generalizing to unseen scenarios than models that only map inputs directly to actions. _Developers evaluating robotics architectures can track emerging world-action model approaches through daily.dev._ ### How much training data does a robot need to learn a new task after pretraining on human video? A robot can learn new tasks with as little as 45 minutes of demonstrations, and in some cases just tens of minutes, after being pretrained on a general foundation model built from over one million hours of human egocentric video. The pretraining gives the model strong baseline physical knowledge, so only minimal task-specific fine-tuning with a handful of demonstrations is needed to adapt it to new tasks like folding napkins or retrieving items from a fridge. _Those weighing pretraining versus fine-tuning strategies for embodied AI follow this kind of scaling detail on daily.dev._ ### Why do dexterous robot hands overheat and limit general-purpose robot capabilities? Current dexterous robot hands lack durability and overheat quickly during extended use, which bottlenecks hardware stability needed to train and evaluate models on complex manipulation tasks. Engineers have had to use ice packs to cool hands during testing. This hardware limitation, combined with missing sensory modalities like touch, hearing, and smell that humans rely on, slows progress toward tasks beyond simple grippers, such as setting up a studio or handling complex multi-step workflows. _Engineers tracking hardware bottlenecks in physical AI deployments can follow developments like this on daily.dev._ --- Tags: [#robotics](https://daily.dev/tags/robotics) [View this post on daily.dev](https://daily.dev/posts/dyna-2-general-purpose-physical-agents-with-hour-long-autonomy-jason-ma-dyna-robotics-sittjfsm7) ```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":"Dyna-2: General Purpose Physical Agents with Hour-Long Autonomy — Jason Ma, Dyna Robotics","url":"https://daily.dev/posts/dyna-2-general-purpose-physical-agents-with-hour-long-autonomy-jason-ma-dyna-robotics-sittjfsm7","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/dyna-2-general-purpose-physical-agents-with-hour-long-autonomy-jason-ma-dyna-robotics-sittjfsm7"},"datePublished":"2026-10-13T22:30:40.069Z","dateModified":"2026-10-13T22:39:51.357Z","description":"Dyna Robotics co-founder and Chief Scientist Jason Ma discusses Dyna-2, a world-action model pre-trained on over one million hours of human video, and the...","image":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/c493b772ab412cbc3e775dd2244e1c88?_a=AQAEuop","thumbnailUrl":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/c493b772ab412cbc3e775dd2244e1c88?_a=AQAEuop","isAccessibleForFree":true,"articleSection":"Latent Space","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":"Latent Space","logo":"https://media.daily.dev/image/upload/s--DnJ9laFj--/f_auto,q_auto/v1780213196/logos/latentspace?_a=BAMAMiWQ0","url":"https://daily.dev/sources/latentspace"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/dyna-2-general-purpose-physical-agents-with-hour-long-autonomy-jason-ma-dyna-robotics-sittjfsm7","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":0},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"robotics","timeRequired":"PT73M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Latent Space","item":"https://daily.dev/sources/latentspace"},{"@type":"ListItem","position":3,"name":"Dyna-2: General Purpose Physical Agents with Hour-Long Autonomy — Jason Ma, Dyna Robotics"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/dyna-2-general-purpose-physical-agents-with-hour-long-autonomy-jason-ma-dyna-robotics-sittjfsm7#faq","mainEntity":[{"@type":"Question","name":"What is a world-action model in robotics and how does it differ from a vision-language-action model?","acceptedAnswer":{"@type":"Answer","text":"A world-action model predicts future video or state of the physical world and uses that prediction to generate robot actions, similar to a chain-of-thought reasoning process in visual and physics space before acting. Dyna-2 is built on a diffusion transformer architecture trained on over one million hours of human video, combining future video prediction with robot action generation, making it more dexterous and better at generalizing to unseen scenarios than models that only map inputs directly to actions. Developers evaluating robotics architectures can track emerging world-action model approaches through daily.dev."}},{"@type":"Question","name":"How much training data does a robot need to learn a new task after pretraining on human video?","acceptedAnswer":{"@type":"Answer","text":"A robot can learn new tasks with as little as 45 minutes of demonstrations, and in some cases just tens of minutes, after being pretrained on a general foundation model built from over one million hours of human egocentric video. The pretraining gives the model strong baseline physical knowledge, so only minimal task-specific fine-tuning with a handful of demonstrations is needed to adapt it to new tasks like folding napkins or retrieving items from a fridge. Those weighing pretraining versus fine-tuning strategies for embodied AI follow this kind of scaling detail on daily.dev."}},{"@type":"Question","name":"Why do dexterous robot hands overheat and limit general-purpose robot capabilities?","acceptedAnswer":{"@type":"Answer","text":"Current dexterous robot hands lack durability and overheat quickly during extended use, which bottlenecks hardware stability needed to train and evaluate models on complex manipulation tasks. Engineers have had to use ice packs to cool hands during testing. This hardware limitation, combined with missing sensory modalities like touch, hearing, and smell that humans rely on, slows progress toward tasks beyond simple grippers, such as setting up a studio or handling complex multi-step workflows. Engineers tracking hardware bottlenecks in physical AI deployments can follow developments like this on daily.dev."}}]} ```