Jason Ma thinks a robot does its best work when it first pictures what is about to happen. Ma, co-founder and chief scientist of Dyna Robotics, explained on the Latent Space podcast why his company builds what it calls a world-action model instead of the approaches more common in the field. He also described the discovery that would prove him wrong. The episode was published October 13, 2026.
What goes in, and what comes out
Ma started with the inputs that he said any robot model needs. The most important is video: images from the robot's cameras covering the past several seconds. A robot also knows the angles of its own joints, from sensors called joint encoders, and Ma said that reading is something a pure video model never sees. The third input is language, so a person can tell the robot which task to do.
The basic output is a set of actions the robot's hardware can carry out. A world-action model adds a step before that. It can predict the future video, "or some representation of it," and then use that prediction to guide its actions. Ma compared this to chain of thought, the step-by-step working that reasoning chatbots produce before giving an answer. In a world-action model, he said, "the models are reasoning in the visual, the physics space, and then outputting actions."
He based the design on how people move. When we act, Ma said, our brains are "implicitly thinking about how the world will evolve" as a result. He said a model built the same way turns out more dexterous and more robust, and better at handling environments it has never seen. He described Dyna-2 as the first world-action model trained on more than one million hours of data.
Dyna's August 2026 technical report adds a qualification. In the report's scaling study, the model learns by training to predict video, but at run time it stays reactive: it neither generates predicted future video nor draws on it.
From noise to smooth motion
Under the hood, Ma said, Dyna-2 is a diffusion transformer, and he explained why each half matters. The transformer, the architecture behind language models, "can just ingest the most amount of data." Diffusion handles precision. A diffusion model starts from random noise and refines it step by step into something detailed. In image generation that detail means sharp edges and clean color contrast. For a robot, it means movements that must be precise and smooth. Ma said the whole field, not only Dyna, has adopted this combination for robot foundation models. The paper that introduced diffusion transformers found that giving the transformer more computation improved the quality of generated images.
Dyna trains the model with flow matching. Ma called it a special case of diffusion that is simpler to tune, because engineers no longer have to hand-adjust the schedule for turning noise into a finished output. He said most image and video generation moved from diffusion to flow matching about two to three years ago. The original flow-matching paper describes the method as learning a velocity field: a set of directions that carries random noise toward realistic data.
The conversation also covered why the method fits robots. A generated image comes out once, but a robot controller can re-evaluate in real time, and Ma agreed with that point. He described the scale of the job: roughly a hundred continuous action dimensions, and a model predicting a vector of at least 50 dimensions 10 to 20 times a second to produce continuous motion. A bigger action space makes it harder to fit the model to the data, he said, and Dyna worked to get a smooth result. Past context matters too. Halfway through opening a microwave, he said, a single snapshot cannot tell you whether the door is opening or closing.
Not the industry standard
Ma said world-action models are "not standard." Some companies, including Runway, are adapting video world models for robots. Runway's GWM-1 announcement describes a robotics variant that generates video conditioned on robot actions. Runway says it could serve as a backbone for robot policies, a source of synthetic training data and a way to test policies on generated scenarios. Other companies build vision-language-action models, or VLAs, which adapt AI models that already understand images and text so that they also output robot commands.
Physical Intelligence's π0 is a well-known VLA. It pairs a pretrained vision-language model with a separate "action expert" that uses flow matching to generate chunks of 50 continuous actions. So the two camps share much of their machinery. The difference Ma stressed is what the model is asked to predict.
His argument starts from the idea that any physical system is about deciding what to do next from what it knows now. A robot is part of the world. It changes its surroundings, and its surroundings shape how it should behave. "The most first principle thing you can do is just predict as much of the world that you can possibly predict," he said. A VLA predicts only actions, or a more abstract summary of the world, and Ma said that "whenever you're predicting abstract representation of the world you're encoding your own bias" about what matters.
Here he invoked the Bitter Lesson. In that 2019 essay, the computer scientist Richard Sutton argued that general methods able to use ever more computation eventually beat systems built around hand-coded human knowledge. Sutton's examples were chess, Go, speech recognition and computer vision. Ma's version for robots is to remove human biases, take the rawest representation, the camera inputs, and predict all of it.
He said robotics will scale like other areas of machine learning, but finding the right data is harder. In language the data is text. In robotics, the sensors a company chooses decide what data it gets. He was skeptical of research that builds in specific assumptions, such as making a robot draw boxes around objects to improve its perception. The lesson, he said, is to "keep the algorithm simpler and scale the right kind of data." Robotics has no equivalent of the internet to draw on, he added, so companies must build the hardware and sensors before they can scale at all.
Where he could be wrong
Asked what the most likely alternative would be if his approach turned out to be wrong, Ma pointed to a gap in his own argument. Predicting as much of the world as possible only works if the sensors capture what matters. A model trained purely on video could be very powerful, he said, yet still "may not be a sufficiently rich enough world model" because it misses sound and touch.
His failure scenario was hypothetical. A rival might release a model that combines vision with touch and turns out to be "a hundred x more data efficient." Vision is highly redundant, Ma explained: neighboring pixels barely differ, and people can understand the world without perfect eyesight. Touch and other senses may carry far denser information, so predicting them could reach the same performance with much less data and computing power.
Even in that case, Ma argued, robotics companies would build such a model better than video-generation companies moving into robotics. Understanding sensing is complicated, he said, and Dyna has worked on it since the company started: "we know the bottleneck, we know how to combine them." A company that only adapts a video model to output robot actions, he said, would lack the know-how to make touch work well.
The question of senses came up once more when heat was raised. Infrared, Ma said, could be just another color, another input to the model. He said he knows companies that use infrared in cooking to judge how well food is cooked, something that is hard to tell from an ordinary color camera.