AivexaNewsSearch
AI news for builders and product teamsChecked every hour

Dyna-2: General Purpose Physical Agents with Hour-Long Autonomy — Jason Ma, Dyna Robotics

Collected Oct 13, 2026

Dyna Robotics has followed its Dyna-2 world-action model with Dyna-2.1, a physical robot agent the company says can sustain hour-long autonomous workflows. Co-founder and Chief Scientist Jason Ma described the release and its underlying architecture in an on-site interview with the Latent Space podcast, published alongside a live demo of the company's Taku robot.

Dyna-2 was announced last month as what Ma calls the first world-action model trained on more than one million hours of data, most of it human video. The company claims the model exhibits several new scaling laws on human data spanning four orders of magnitude. As a world-action model, it predicts future video frames, then uses that prediction to generate the actions a robot executes. Ma frames this as analogous to chain-of-thought reasoning: the model reasons in visual and physical space before committing to motion. The payoff, he says, is greater dexterity, robustness, and generalization to unseen environments and scenarios.

On inputs and outputs, the model consumes video from the robot's cameras covering the past several seconds, joint angles from the robot's actuators and encoders, and language instructions. Most robot foundation models take only video and language; the proprioceptive signals are a distinguishing feature. Outputs include executable actions and predicted future video, hence the "world-action" label.

Under the hood, Dyna-2 is a diffusion transformer with an encoder-only transformer backbone. Ma explains the choice from first principles: transformers ingest the most data, while diffusion refines noisy samples into precise, smooth, fine-grained motion, the same property that makes diffusion good at generating sharp images. The company uses flow matching, a simpler-to-tune variant of diffusion that reduces the number of iterative steps required; the podcast conversation cites a reduction from roughly 40 to 100 diffusion steps down to about four. Flow matching has largely displaced plain diffusion in image and video generation over the past two to three years.

The adaptation story is central to Dyna's commercial pitch. A single generalist base model can be fine-tuned to new tasks from as little as 45 minutes of demonstrations, according to Ma. The office demo included a laundry-folding robot adapting to an unfamiliar towel and a robot pulling drinks from a fridge. The base model supplies broad physical priors, and light fine-tuning makes it an expert at a specific job.

Dyna's data strategy leans on egocentric human video because robotics has no equivalent of the internet, Ma argues, so no single company can collect enough on its own. He compares a lone dataset to trying to build the internet with only Reddit or Facebook. Human video covers a vast range of physical activity and gives the model a strong prior, but it leaves a gap between human and robot actions that must be closed with robot-collected demonstrations, which are more expensive to gather. Ma describes data quality as non-stationary: what counts as high quality changes as the model improves. Dyna's Dyna-2 blog post invites data vendors to reach out via an email address listed there.

Other points from the conversation: Dyna builds its own wheeled semi-humanoid hardware rather than buying off-the-shelf bipedal robots, and Ma argues robotics needs on-device inference, model distillation, and custom hardware. Dexterous hands that overheat, along with missing sensory data, remain among the field's hard bottlenecks. Agentic harnesses let the robots chain multi-step physical workflows over hours.

Deployment results are the headline numbers. Ma says Dyna's models reached a 99.4% success rate folding restaurant napkins for 24 hours straight, including recovering after pulling over the whole stack. The company says its robots fold thousands of napkins daily at Din Tai Fung and has deployments at Best Western, which Ma contrasts with the many robotics pilots that never scale.

A likely trade-off worth noting is that a 99.4% success rate over 24 hours still leaves room for roughly one failure every few hundred attempts, and recovery behavior, not raw reliability, is what keeps a long-horizon run going. This reporter's reading is that Dyna's emphasis on napkin folding and drink retrieval reflects a deliberate choice to compete on sustained reliability in narrow commercial tasks rather than on general-purpose demos, which is where most robotics startups have stalled.

Why it matters: For developers and product teams working on embodied AI, Dyna-2.1 is a concrete data point that world-action architectures plus flow matching are being pushed into production, not just research. The claims of 45-minute task adaptation and 24-hour autonomy suggest the competitive frontier is shifting from model capability to reliability, deployment, and data pipelines. Inference costs, hardware constraints, and hand durability remain the limiting factors, so teams evaluating similar systems should weigh sustained success rates and recovery behavior over peak benchmark numbers.

Read at Latent Space

Based on reporting from the original publisher. Visit the source for full context and later updates.

Publisher excerpt

Our first robotics pod with Dyna’s Chief Scientist on their big bet on World Action Models & the Next Frontier of Physical AI