By playing, you agree to uListen's Terms and Privacy Policy and YouTube's Terms of Service.

Robots That Predict the Future: Dyna-2 & Physical AI — Jason Ma, Dyna Robotics

From training a world-action model on over one million hours of human video to deploying robots that fold thousands of napkins daily at Din Tai Fung, Dyna Robotics is pushing physical AI beyond impressive demos and into real-world production. In this episode, Dyna Robotics co-founder and Chief Scientist Jason Ma joins swyx on-site to unpack how robots can learn new tasks from as little as 45 minutes of demonstrations, why predicting the future of the physical world may be the key to general-purpose robotics, and what it actually takes to build robots that work reliably for real customers. We go deep on Dyna-2, Dyna's million-hour world-action model that combines future video prediction with robot actions, the scaling laws behind human-to-robot transfer, diffusion transformers and flow matching, and why synthetic environments and self-generated training data could unlock the next wave of robotics progress. Jason also explains why Dyna builds its own hardware instead of relying on off-the-shelf humanoids, how agentic harnesses enable robots to execute hour-long workflows, why dexterous hands and missing sensory data remain major bottlenecks, and what separates real commercial deployments from the graveyard of robotics startups. We also cover world-action models versus vision-language-action models, GPT-6 and coding agents in robotics, and Jason's journey from AlphaGo and Penn to NVIDIA, Google DeepMind, and founding Dyna. We discuss: - Why Dyna-2 was trained on over one million hours of human video - World-action models: predicting future video to guide physical robot actions - Diffusion transformers, flow matching, and real-time robotic control - How robots learn new tasks with as little as 45 minutes of demonstrations - Why robotics has no equivalent of the internet and data remains its biggest scaling challenge - Human egocentric video, robot demonstrations, and cross-embodiment transfer - Synthetic environments, simulation, and models that generate their own training data - Why robots need on-device inference, model distillation, and custom hardware - Why Dyna chose wheeled semi-humanoids instead of traditional bipedal robots - The hardware bottleneck: why dexterous robot hands still overheat - Best Western, Din Tai Fung, and the difference between robotics pilots and scaled deployments - The Bitter Lesson of Robotics: scaling data instead of engineering specialized solutions - Agentic harnesses and robots executing multi-step, hour-long physical workflows - GPT-6, coding agents, and AI-generated simulation environments for robotics research - World-action models versus VLAs, and why vision alone may not be enough - Jason's journey through AlphaGo, Penn, NVIDIA, Google DeepMind, and the Eureka project - Dyna's research culture, deployment-first philosophy, and vision for millions of robots --- Sponsorship and business inquiries: business@latent.space --- Jason Ma - LinkedIn: https://www.linkedin.com/in/jason-ma-742224a2/ - X: https://x.com/JasonMa2020 Dyna Robotics - Website: https://www.dyna.co/ - X: https://x.com/DynaRobotics Timestamps 00:00:00 Hook 00:01:56 Introduction and Dyna-2's World-Action Model 00:05:04 Diffusion Transformers, Flow Matching, and Generalist Robots 00:10:15 The Robotics Data Bottleneck 00:15:46 Synthetic Data and Self-Improving World Models 00:18:47 Real-Time Inference and Model Distillation 00:21:41 Full-Stack Robotics: Hardware and Supply Chains 00:26:58 Why Dyna Built a Wheeled Semi-Humanoid 00:31:31 Teaching Robots New Tasks and Evaluating Performance 00:33:55 Autonomous Workflows and Hardware Limitations 00:38:02 Real-World Deployments at Best Western and Din Tai Fung 00:43:03 The Bitter Lesson of Robotics 00:48:09 Agentic Harnesses for Physical AI 00:51:03 GPT-6, Coding Agents, and AI-Native Robotics 00:53:58 World-Action Models vs. VLAs 01:01:13 From AlphaGo to Founding Dyna Robotics 01:06:16 Dyna's Research Culture and the Future of Robotics

Jason MaguestswyxhostAlessio Fanellihost

Oct 13, 20261h 21mWatch on YouTube ↗

Summaries, chapters, transcripts, and FAQs on this page are made by uListen, not YouTube.

WHAT IT’S REALLY ABOUT

Dyna Robotics’ world-action model for fast, deployable generalist robots

Jason Ma explains Dyna-2 as a world-action model that predicts future visual world states and outputs robot actions to improve robustness and generalization.
The system is trained primarily on large-scale video-based data (including egocentric human data) plus smaller amounts of robot data, with the company actively seeking more diverse and task-specific data sources.
Dyna uses diffusion transformers (and flow-matching variants) to produce smooth, precise continuous robot control, and says context/history is critical for disambiguating actions in physical tasks.
He argues robotics differs from language models mainly in real-time constraints, requiring fast on-device inference, full-stack optimization, and often distillation from larger trained models.
The episode highlights Dyna’s product-driven, deployment-first approach, including scaled napkin-folding deployments at Din Tai Fung and laundry-related work for hotel customers like Best Western.

IDEAS WORTH REMEMBERING

5 ideas

Dyna’s core bet is a “world-action model” that predicts futures, not just actions.

He describes it as predicting future video/world state in addition to producing executable robot actions, arguing this helps dexterity, robustness, and generalization because the model can “think through” how the world evolves under actions.

Inputs blend video, robot joint state, and language; outputs include both actions and predicted future observations.

The model ingests robot camera video plus robot proprioception (joint encoders) and can also take language instructions; it outputs robot actions and can also generate future video (or a representation of it) that then guides action selection.

Diffusion/flow-matching is used to get the smooth, precise motion robotics requires.

Jason Ma says Dyna-2 is a diffusion-transformer-based model (with flow-matching as a simpler-to-tune diffusion variant) because robotics actions need smooth, precise, fine-grained control analogous to high-quality image generation.

Robotics’ data bottleneck drives a hybrid strategy: scale human egocentric data, then adapt with robot data.

He frames robotics data as the biggest constraint because “there is no internet,” and says “high-quality” data shifts over time; Dyna combines large-scale egocentric human data to learn broad physical priors with smaller, expensive robot-collected data for adaptation.

Task onboarding is presented as fast: minutes-to-hours of teleoperated data plus iterative real-world eval.

Jason Ma says new tasks can be learned with “a couple hours” of robot data collection (sometimes “tens of minutes”), followed by training, real-world rollout, measuring performance on-hardware, and iterating by collecting more data where failures appear.

WORDS WORTH SAVING

5 quotes

World-action model is a model that's able to predict the future video, the future state of the world, so hence the name world model. But in addition, you can also use that future generation to also produce actions robots can execute in the physical world to control the robot.

— Jason Ma

Today there's a lot of robotics companies just talking about deployments, talking about pilots, but what's really hard is not putting a robot in one location or doing a pilot. What's really hard is that the customers believe or validate that your robot creates enough value that they wanna deploy your robots across many of their sites.

— Jason Ma

For robots, like, the inference latency has to be, like, very short, right? The robot needs to act. Like, it can't just hold a knife and not do anything for three second. That's gonna freak people out, right?

— Jason Ma

Robotics is going to scale just like any other domains that machine learning has solved, right? So what does that mean? That means have as much data as possible.

— Jason Ma

Because hardware decision is much more expensive and much more costly, just 'cause hardware iteration is a lot slower, you actually want to take a incremental iterative approach, not going all the way to the final form factor.

— Jason Ma

World-action models and future predictionDiffusion transformers and flow matching for controlRobotics data bottleneck and data acquisitionEgocentric human data vs robot-collected dataSynthetic data and self-improving training flywheelsLow-latency on-device inference and distillationFull-stack robotics: hardware, actuators, and supply chains

High quality AI-generated summary created from speaker-labeled transcript.

Get more out of YouTube videos.

Read along on your phone, search it beside the video, or ask your favorite AI app what was said.