Skip to content
Main Site News Console

Dyna-2: A General-Purpose Physical Agent with Hour-Scale Autonomy — Jason Ma, Dyna Robotics

· Latent Space Translated
播客深度访谈

Be sure to tune in to AIE NYC, and join us in San Francisco next month for our AIE CODE conference!

This episode features a live demo of the Taku robot—we recommend watching on YouTube for the best experience.


Last month, Dyna caused a stir by releasing Dyna-2, a World-Action Model pretrained on over 1 million hours of human video that claimed to uncover multiple new scaling laws spanning four orders of magnitude on human data. Now, they’ve followed it up with their physical robot agent, Dyna-2.1, and the results are stunning:

Dyna Robotics (@DynaRobotics): We are officially releasing Dyna-2.1, the first physical agent capable of reliable, ultra-long-horizon full-body autonomy. Combining our all-new semi-humanoid hardware with an agentic system built around Dyna-2, it can handle extended real-world workflows.

Here is uncut real-world footage…

Dyna-2.1 Demo Preview

In his talk at AIE this year, Jason discussed how the real challenge for real-world robotics is reliable, long-horizon autonomy—and how Dyna achieved a 99.4% success rate with its model folding restaurant napkins continuously for 24 hours, even recovering autonomously after knocking over an entire stack of napkins.

From training World-Action Models on over 1 million hours of human video to deploying robots folding thousands of napkins daily at Din Tai Fung, Dyna Robotics is pushing Physical AI beyond flashy demos and into real-world production. In this episode, Dyna Robotics co-founder and Chief Scientist Jason Ma joins swyx in person to dive deep into how robots can learn new tasks with just 45 minutes of demonstrations, why predicting the physical world’s future might be the key to general-purpose robotics, and what it truly takes to build robots that operate reliably for real customers.

Watch Video: YouTube

We dive deep into Dyna-2 (Dyna’s 1-million-hour World-Action Model combining future video prediction with robot actions), the scaling laws behind human-to-robot transfer, Diffusion Transformers and Flow Matching, and why synthetic environments alongside self-generated training data could unlock the next wave of robotics. Jason also breaks down why Dyna chose custom in-house hardware over off-the-shelf humanoids, how agentic harnesses empower robots to execute multi-hour workflows, why dexterous hands and missing sensory data remain major bottlenecks, and what separates real commercial traction from robotics startups that fail. We also cover World-Action Models vs. Vision-Language-Action Models (VLAs), GPT-6 and coding agents in robotics, and Jason’s journey from AlphaGo and Penn through NVIDIA and Google DeepMind to founding Dyna.


In this episode:

  • Why Dyna-2 was trained on over 1 million hours of human video
  • World-action models: Guiding physical robot actions by predicting future video
  • Diffusion Transformers, Flow Matching, and real-time robot control
  • How robots can learn new tasks from just 45 minutes of demonstrations
  • Why robotics has no equivalent to the internet-scale data deluge, and why data remains its biggest scaling bottleneck
  • Human egocentric video, robot teleoperation, and cross-embodiment transfer
  • Synthetic environments, simulation, and models that generate their own training data
  • Why robotics demands on-device inference, model distillation, and custom in-house hardware
  • Why Dyna chose a wheeled semi-humanoid over traditional bipedal humanoids
  • Hardware bottlenecks: Why dexterous hands still struggle with overheating
  • Best Western, Din Tai Fung, and the difference between robotics pilots and scaled production deployments
  • The Bitter Lesson of Robotics: Scaling data rather than engineering bespoke solutions
  • Agentic harnesses and robots executing multi-step, multi-hour physical workflows
  • GPT-6, Coding Agents, and AI-generated simulation environments for robotics research
  • World-action models vs. VLAs, and why vision alone might not be enough
  • Jason’s journey across AlphaGo, UPenn, NVIDIA, Google DeepMind, and the Eureka project
  • Dyna’s research culture, deployment-first philosophy, and the long-term vision of millions of robots

Jason Ma

Dyna Robotics


Timestamps

00:00:00 Intro & Dyna-2’s World-Action Model
00:03:08 Diffusion Transformers, Flow Matching, and General-Purpose Robotics
00:08:19 The Data Bottleneck in Robotics
00:13:50 Synthetic Data & Self-Evolving World Models
00:16:51 Real-Time Inference & Model Distillation
00:19:45 Full-Stack Robotics: Hardware & Supply Chain
00:25:02 Why Dyna Built a Wheeled Semi-Humanoid Robot

#播客#深度访谈#具身智能#AI智能体#4Allapi.com

Published by the 4All API team

Original link:https://www.latent.space/p/dyna

Need a mainstream LLM API? 4All API gives you one key to call OpenAI, Anthropic, Google Gemini, Qwen, DeepSeek, and dozens more — at official-pass-through pricing, with enterprise-grade reliability, integrated in 5 minutes.

Sign up for the 4All API console →