Robots shouldn't just learn actions—they must learn to predict the whole world first
Dyna's world-action model lets robots generate future video in their imagination first, then generate actions accordingly; the founder calls it chain-of-thought for robots, and it's the fundamental divide with the VLA approach.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
Robots that visualize the future outperform those that memorize actions
Dyna-2 uses a single model that both predicts future video or world state and generates robot actions from it. Jason Ma says this resembles how the human brain implicitly simulates "what happens next if I move this way"—and then acts accordingly. He frames it as chain-of-thought for robotics: the model reasons first in visual and physics space, then outputs motion. This makes robots more dexterous, more robust, and better able to handle scenarios unseen during training.
— Jason MaRobots have no internet: every company must manufacture its own training corpus
Language models leverage massive internet text for pretraining; robots have no equivalent corpus, so each company must find its own data source. Jason Ma notes that internal collection alone falls far short—it's like trying to build "the internet" using only Reddit or Facebook. Dyna's strategy layers human first-person video to cover nearly all physical manipulation tasks, then fine-tunes with scarce, expensive real robot data to bridge the gap between human motion and robot execution.
— Jason MaSynthetic data has crossed a threshold: robots can now generate their own training loops
At NVIDIA, Jason Ma worked on Eureka, which used GPT-4 to auto-generate simulation environments and reward functions for teaching a robotic hand to spin a pen—results were limited because the model lacked genuine visual and 3D understanding. Today's frontier models have CAD and 3D reconstruction capabilities; Ma believes synthetic data has hit an inflection point. Better yet: Dyna's world-action model itself generates video, so the model can produce its own training data, creating a self-reinforcing flywheel.
— Jason MaLatency is non-negotiable: robots must reason on device, not in the cloud
A language model can afford to wait a few seconds; a robot holding a knife cannot freeze for three seconds. Dynamic tasks like tennis demand instant response to gravity and physics. So deployed models must be small enough to run locally and infer in real time—no cloud dependency. This doesn't mean the training model must be small; the approach mirrors what language models pioneered: train as large a model as possible, then distill a smaller one that ships on hardware.
— Jason MaHardware bottlenecks, not algorithms, are the true barrier to scale
Dyna is a full-stack company. Mature components like lidar and cameras are bought off the shelf, but actuators determine how precisely the robot can be modeled—so Dyna builds its own. Rare-earth magnets are controlled by roughly two countries. Even if assembly moved stateside, raw materials still flow through global supply chains; the longer the supply chain, the longer the path from zero to one. Ma argues the industry over-indexes on models and data while drastically underestimating hardware complexity.
— Jason MaWheels beat legs when safety and power efficiency outweigh anthropomorphism
Dyna's robot is semi-humanoid: shoulders, arms, and a four-degree-of-freedom waist up top; wheels below. Ma's reasoning: quadruped robots demand high power, stability is poor, and safety cannot yet be assured—dancing robots on social media prove it by falling constantly. But the waist's twist degree-of-freedom is intentional: human daily motion relies almost entirely on the torso and back; to learn motion from human data, robots need corresponding skeletal degrees of freedom to transfer that knowledge.
— Jason MaScaling to one location is easy; customers scaling you to every location is new
Dyna folds towels at Best Western and napkins at Din Tai Fung, with cutting tasks ahead. Ma stresses that many robotics companies demo at one site; the hard problem is customers validating the value and requesting deployment to every location—unprecedented in general-purpose robotics, he says. Din Tai Fung locations originally folded inconsistently; the robot brought the consistency customers genuinely wanted, driving them to renew contracts and demand cutting capability too.
— Jason MaPixel prediction defeats representation learning when avoiding human bias matters most
Ma contrasts Dyna's architecture with the VLA path: VLA essentially predicts an abstract representation, meaning engineers chose what matters and what gets encoded, baking in human bias. The more first-principles approach: predict raw camera input directly, letting the model decide what's important. He concedes this judgment could be wrong—if touch or sound outweigh vision in importance or data efficiency, pure-vision world models could be superseded. But he expects roboticists to better understand sensor fusion than video-generation teams pivoting into robotics.
— Jason MaIn their own words · checked verbatim
in our brain, we are implicitly thinking about how the world will evolve based on our actions. So now we have a model architecture that’s able to do that to guide robot actions.
Jason Ma0:00
So I think data is the most important thing in robotics today, because unlike language model, there is no internet, right? So every company that’s building model, building hardware has to think about data.
Jason Ma8:19
It’s kind of like humans, you know. If, Sean or me, we go work at Rosewood Hotel tomorrow, we are not gonna be able to make the bed the way they want, right? We need a lot of training.
Jason Ma12:26
The robot needs to act. Like, it can’t just hold a knife and not do anything for three seconds. That’s gonna freak people out, right?
Jason Ma17:38
So our robots today fold thousands of napkins for many different locations, many different Din Tai Fung locations in the United States. So that’s actually our biggest scaled customer, right?
Jason Ma37:09
And the bigger lesson tells you, let’s remove human biases. Let’s just take the most raw representation, the camera inputs, and just predict all of that right?
Jason Ma54:07
The search space is larger than the total number of atoms in the world. How could you possibly search for solutions in that?
Jason Ma59:17
Figures
| Dyna-2 training data | Over 1 million hours | 0:00 |
| Full laundry cycle time | ~1 hour | 33:06 |
| Cutting task training data | ~45 minutes | 29:35 |
| Next year deployment target | Hundreds of robots | 51:18 |
| Hardware team ratio / Model team size | ~30-40% / ~10 people | 1:06:44 |
Glossary
- World Action Model
- A model architecture that both predicts future video or world state and outputs robot actions from those predictions.
- Diffusion Transformer
- A transformer-based model that uses a diffusion process to gradually refine noisy predictions into precise actions.
- Flow Matching
- A simplified diffusion variant with simpler scheduling that generates actions in fewer steps.
- VLA (Vision-Language-Action model)
- An adapted vision-language model that directly outputs robot actions from visual and language inputs.
- Eureka
- A research project that uses large language models to automatically generate simulation environments and reward functions for robot training.
- Distillation
- Training a large model first, then compressing it into a smaller model that runs locally on robot hardware in real time.
How to listen
Engineers and investors tracking embodied AI and robotics startups; anyone wanting to understand the world-model architecture debate, data strategy, and full-stack hardware approaches.
Information density is uniform throughout; if time is tight, skip the closing remarks about the company name's origin.