Language models had the internet to learn from. Robots have nothing like it, and Jason Ma thinks that fact shapes nearly every choice a robotics company makes. "Data is the most important thing in robotics today because unlike language model there is no internet," Ma, co-founder and chief scientist of Dyna Robotics, said on Latent Space. In the episode, published October 13, 2026, he described the company's answer: a large base of human video, a small amount of data recorded on the robot itself, and a growing interest in data that machines generate.
Why one company's data isn't enough
Dyna collects data itself, and robots are "everywhere in the office," Ma said. But he argued that no single company can collect enough on its own. Trying is "kind of like you're trying to build an internet by just having Reddit or just having Facebook." Those are great websites with plenty of users, "but it's not enough." So some data has to be acquired: bought, or found among what already exists.
He also called data quality in robotics "a non-stationary problem": what counts as high-quality data for a model changes over time. A model that knows nothing can learn from almost anything, including a lot of ordinary video. As it improves, the useful data gets more specific.
The broadest source is egocentric video, footage filmed from a person's own point of view while they do things. Ma called it "a very good way of pre-training robots" because human activity is so varied that it covers "every single physical activity you can possibly imagine." Pretraining is the first, broad phase of training, before a model is tuned for particular jobs. Human video, he said, gives the model a strong prior, a built-in expectation, about how the world changes around a body acting in it.
The catch is the body. Until a robot looks exactly like a human, he said, there will be a gap between human movements and robot movements, so a model trained only on people can't simply drive a robot. It still needs data recorded on the robot, which is "much more expensive and much harder to collect." His recipe is to use the largest source available, egocentric video, for the broad picture of physical manipulation, then add a small amount of robot data to master a specific task. Human data covers one part of robot cognition, he said, while robot data covers low-level reflex and dexterity. There is no single winning type, in his view: both have to go into one model.
Getting outside recordings into usable shape is its own job. Dyna's infrastructure account describes syncing camera footage with the robot's joint readings and rejecting episodes with blacked-out cameras, discontinuous joint data, missing or hidden hands, or damaged frames. The company says it raised its weekly intake from 14,000 to 440,000 episode-hours.
What a million hours bought
Dyna-2, the model Dyna described in an August 2026 technical report, was trained on more than a million hours of human data as its starting point, Ma said. From there, he said, the model can take on many different tasks with "a very minimal amount of fine tuning": tell it the task and give it a few demonstrations.
The report points the same direction, with conditions. Dyna fine-tuned models for 14 tasks on a real robot using at most ten hours of demonstrations per task. As human pretraining grew from 1,000 to 10,000, 100,000 and one million hours, the average score rose from 20% to 28%, 45% and 53%. Those scores give partial credit and are normalized to each task's ceiling, so they are not plain success rates. One task, opening a bottle cap, used about ten minutes of demonstrations and succeeded half the time at the largest scale.
Ma still called a million hours "just the beginning." Dyna wants more raw, varied footage, but also custom data for narrow tasks, collected in-house or by partners. He compared robots to people starting a new job: if he or the host went to work at the Rosewood Hotel tomorrow, they would not be able to make the bed the way the hotel wants without a lot of training. Specialized data of that kind, similar to what Mercor supplies for language models today, will be very valuable for robotics too, he said. He also pointed data vendors to a data-partnerships email address at the end of the Dyna-2 blog post.
A couple of hours per task
To teach a new task, Ma said, Dyna collects a little robot data, which he put at "a couple hours," or "just tens of minutes" for some tasks. The discussion cited a 45-minute example. Collection here means teleoperation: a person remote-controls the robot through the task while its movements are recorded. Offered the idea of a new task as an onboarding challenge for engineers, Ma said: "pitch me a new task and i'll come to the office you know collect some data train a model and see what happens."
Testing happens in the real world. After training, Dyna runs the model on the same setup where the data was collected. Sometimes it works well right away. When it is at "maybe like 70 percent 80 percent," the team looks at what it gets wrong and collects more data. Ma likened this to reinforcement learning from human feedback, the method used to tune chatbots with human ratings: the robot attempts the task, someone scores it good, bad or somewhat good, and the robot's own experience becomes training data. Eventually, he said, "anyone should be able to train tasks on it."
Training before the hardware existed
Ma offered Dyna's new robot as evidence that human pretraining transfers to new machines. The robot was ready "not so long ago," so Dyna has much less data for it than for its stationary robot arms, and in some pretrained models it was never seen at all. Even so, he said, the model learned the tasks shown very quickly. He credited an early bet: "even before the hardware is ready we already committed to pre-training on humans because we knew at some point our robots will be like human." Training could get ahead while the hardware caught up, and "now we're seeing that like the transfer is actually there."
The next dimension: synthetic data
Asked about building simulated environments on the fly, Ma called synthetic data and simulation "one of the next frontiers." Egocentric and robot data are already scaling, he said, and "synthetic data will be a new dimension of scaling."
He has tried this before. His PhD work focused on training robots in simulation, which he said wasn't ready then. At NVIDIA he worked on Eureka, which he described as the first paper to use a large language model, GPT-4, to generate environments and reward functions for reinforcement learning. In reinforcement learning a robot learns by trial and error, and a reward function is the scoring rule that tells it how well it did. The idea grew from a pen-spinning project. Ma said writing all the reward terms for a simulated robot hand spinning a pen himself would be impossible, so he wondered whether a model that writes code could do it. In his opinion, the result was the first recursive self-improving robotic agent in simulation.
The Eureka paper is narrower than "generating environments" suggests. GPT-4 received existing simulator code and a task description, wrote candidate reward programs, and revised them based on training results. The authors report that its rewards beat human-written ones on 83% of tasks. Nearly all the evidence comes from simulation, and pen spinning also needed a pretraining step and a gradual curriculum. Ma said the approach was held back by the models of the time, which lacked good vision and an innate 3D understanding, so they were basically generating code with limited text feedback.
Now, he said, frontier models' vision is "crazy." They can do real-to-sim, CAD work and 3D reconstruction, putting the technology at a tipping point. Dyna has started using frontier models to generate synthetic environments tailored to its own tasks. What excites him about GPT-6, Ma said, is its ability to construct virtual environments that turn into simulation. OpenAI's announcement of GPT-6 Astra includes demonstrations in Blender, Unreal Engine and mechanical design.
Dyna is also trying a second route. Its world-action model predicts future video as well as robot actions, so, Ma said, it can generate its own training data, which he called a self-fulfilling flywheel. Others are exploring similar ideas: the conversation mentioned Runway's robotics world models, and Runway's GWM-1 announcement lists synthetic training data as one intended use, a capability the company describes rather than one it has proven.
Ma said he set aside many of these ideas when Dyna started at the end of 2024, because building a company meant being practical and simulation wasn't ready. "Now with you know like the new frontier models and what we're seeing internally like this might change," he said.