How Dyna Teaches Robots to Work Beyond a Demo

A general model learns a particular job

Imagine teaching a robot to chop food. Dyna Robotics starts with a model trained on a much broader collection of human activity, then collects demonstrations on the robot for the task at hand. In the Latent Space conversation, co-founder and chief scientist Jason Ma describes a chopping example for which the hosts say 45 minutes of data were collected. Ma says some tasks take tens of minutes of robot data and others a couple of hours.

The distinction matters. Human video provides a broad physical prior; robot data connects that prior to a particular body and task. A person’s hand and a robot gripper do not produce the same actions. Ma says that gap makes direct execution from human video alone unlikely today, even if the video helps a model learn how the world changes around an actor.

Dyna’s Dyna-2 model was trained on more than one million hours of human data, according to Ma. He calls that a starting point, not enough data to finish the job: the company wants more diverse data as well as custom examples for specific tasks. His analogy is a person who can make a bed but still needs training to meet a hotel’s particular standard.

Predicting what the world will do

A world-action model, in Ma’s account, combines two outputs: a prediction of future video or a representation of it, and actions the robot can execute. Its inputs can include recent camera images, joint angles from encoders, and a language instruction. The predicted future is meant to guide action, giving the system a way to reason visually about how the scene may evolve.

For example, an instruction to retrieve a drink requires the robot to act in a changing scene, not merely recognize a fridge. Ma argues that forecasting the world around the robot is useful because the robot affects its surroundings and those surroundings affect what it should do next. He contrasts Dyna’s approach with vision-language-action models, which he describes as predicting actions or a more abstract world representation.

The model also needs history. Ma’s example is a microwave door halfway through a movement: a single observation may not reveal whether the robot is opening or closing it. Recent context helps disambiguate what action belongs next.

Under the hood, Ma describes Dyna-2 as an encoder-only diffusion transformer. He explains diffusion as an iterative transformation from noise to a detailed output: in this setting, the desired output can be precise, smooth motion. He describes flow matching as a related approach that simplifies tuning the transformation.

From demonstration to a working skill

The adaptation loop is concrete: collect robot demonstrations, train, run the model on the real setup, and inspect failures. Ma says that when performance is only partway there, the team can collect more data and iterate; robot experience can also be scored and used for training. Training metrics alone cannot replace evaluation on the physical task.

Ma says human pretraining also helped transfer to a newer robot for which Dyna had less data than for its stationary arms. He credits the transfer to pretraining before the hardware was ready.

More data is not automatically the right data. Ma says a model may miss information needed for a task even after millions of hours if training captures only vision and motion while the task depends on other senses. He suggests that touch or sound might make a model more data-efficient than vision alone.

The body and the clock constrain the model

A robot cannot always wait for a cloud response. Ma gives the example of a robot juggling a ball: gravity keeps acting while inference runs. He says Dyna builds its own inference stack and that a deployed model must be small enough to serve locally. He expects robotics to converge on training larger models and distilling them into smaller ones.

Dyna’s robot has a humanlike upper body and wheels. Ma says wheels offer a practical tradeoff in energy and stability; a movable torso allows movements that a purely vertical lift cannot easily reproduce. The form also matters to the learning strategy: he says the company considered human proportions and torso movement because it pretrains on human demonstrations.

Even with a capable model, hardware can slow learning. Ma says current dexterous hands can overheat and lack durability; that restricts both the tasks a robot can attempt and the speed of physical evaluation. For the laundry workflow he describes, a parallel-jaw gripper and available sensing are sufficient, in his assessment.

A workflow is longer than a motion

A laundry job illustrates why individual motor skills are not the whole system. Ma says a robot needs higher-level decisions: while the washer runs, it can work elsewhere; when the wash finishes, moving towels to the dryer promptly avoids a bottleneck. He describes a hierarchical harness around the world-action model to compose skills and manage those decisions.

Ma says the laundry demonstration completes a full loop on the order of an hour. He presents hours-long unsupervised workflows as the value Dyna is building toward.

The customer test

Ma says Dyna folds laundry towels for Best Western and that its robots fold thousands of napkins per day across multiple Din Tai Fung locations in the United States. The restaurant’s fold has a precise requirement: the seam must leave a pocket for chopsticks.

Cloth makes a useful test because its shape changes as it is handled. Ma argues that a general model adapted to a use case can reach production quality faster than a specialized system. He also says Din Tai Fung is considering food cutting as another use case.

Ma distinguishes a pilot at one location from a customer choosing to deploy across sites. He reports tens of robots deployed and describes hundreds in the following year as a goal, not an achieved count. On this account, deployment is a test of the complete service: the model, physical hardware and customer’s actual work all have to fit together.

What comes after the captured data

Ma sees synthetic environments as a possible additional source of training data. Dyna has explored using frontier models to generate environments for its own settings, and he says its world-action model can generate video that may become training data. He is optimistic that those ingredients could support a self-improving loop.

That interest connects to his earlier research. Ma says a pen-spinning project led him to simulation because a suitable physical hand was unavailable; specifying a reinforcement-learning reward was difficult, so he explored having a code-writing model generate reward functions. He identifies that work as the Eureka project and describes it as a team effort.

Ma says he started Dyna with Lindon Gao and York Yang at the end of 2024 because he wanted robotics research tested in real use. His closing appeal is for more work that integrates foundation models and hardware at customer sites. The practical challenge is specific: learn broad physical behavior, adapt it to a particular body and task, and evaluate whether the resulting system can repeatedly do useful work.