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

Full analysis

Dyna Robotics put a robot that folds napkins into multiple Din Tai Fung restaurants, and it folds thousands a day. Not a demo. A paying customer who signed an expanded contract to add vegetable cutting. The model behind it, Dyna-2, was trained on over a million hours of human video and predicts what the world will look like next before it moves. Jason Ma, the co-founder, says the harness on top runs hour-long jobs like a full laundry cycle with no human watching.

This is a robotics story, not a frontier chatbot story. But it touches the AI stack most readers already pay for: OpenAI's GPT-6 is generating the fake training worlds, Anthropic's Opus 5.5 is in the coding loop, and the whole thing runs on the same distillation and on-device inference problems that shape what your AI tools cost. Hard to undo for nobody here. Nobody has to decide anything today. The question is whether physical AI is now real enough to plan around.

The Skeptic

Napkins. Let me say it plainly: the proof point is folding napkins and towels, and cutting vegetables is the stretch goal. These are flat, forgiving, repetitive tasks in a controlled kitchen. That is exactly where a robot should win first, and it tells you almost nothing about the hour-long laundry claim, which came with no customer attached. Ma admits his own thesis might be wrong, that touch and sound could make a rival 100 times more data-efficient. The hands overheat and need ice packs during evaluation. Read that again. The bottleneck is not in the model. The hardware physically cannot run long enough to test the thing they are selling.

The Researcher

The architecture claim is the interesting part, and it is a real bet, not marketing. Dyna-2 predicts future video frames and uses that prediction to guide its actions. Ma calls it visual chain-of-thought, meaning the model imagines the next second of the world the way a reasoning model writes out steps before answering. The rival approach, VLAs, bolts robot controls onto a language-vision model. Ma's argument is the bitter lesson: raw scale beats hand-built assumptions. Fair. But a million hours is tiny next to the internet, and he says so himself. The flow-matching trick cuts inference from 40 to 100 steps down to about four, and that reduction is what makes real-time control possible at all.

The Compute Pragmatist

The constraint that matters is 50 milliseconds. A robot grabbing a towel cannot wait for a cloud round-trip, so Dyna built its own inference stack and plans to distill the big trained model into a small on-device one. That is the same move your LLM vendors make to cut costs, except the real-time deadline is brutal and the penalty for missing it is a dropped plate, not a slow response. This is why the Cerebras CEO's "AI appliance in every home" line matters: capable-enough models are collapsing onto hardware you own and run locally. Good for a fixed restaurant task. The economics only work because napkin-folding is the same motion a billion times. Novel environments reset that math.

The Open-Source Advocate

Here is the part nobody says out loud. There is no internet for robotics. LLMs got a free web-scale corpus; robots get nothing, so every team has to generate its own data from its own hardware. That kills the open ecosystem advantage before it starts. You cannot download Dyna's million hours the way you pull a Llama checkpoint. Ma is publicly begging for data partnerships at an email address, which tells you how scarce the raw material is. Whoever accumulates proprietary physical data and the robots to collect it owns a position that open weights cannot route around. This is the opposite of the LLM world, where open models sit within 80% of the closed ones. In robotics, the data moat is the whole game.

The Builder

What would I actually deploy on Tuesday? Nothing, unless I run restaurants or hotels. The deployment-versus-demo gap is the useful lesson here, and it cuts against every humanoid-robot video you have seen. Ma's dig at Optimus is pointed: they went through many iterations and scrapped designs because they spent years building for a warehouse instead of shipping a wheeled half-humanoid to a real customer and taking the feedback. That is the same discipline that separates a working AI feature from a slick demo. Ship the narrow thing, let the customer break it, expand from there. Din Tai Fung going from napkins to cutting is that loop working.

Where they disagree

The Researcher thinks the architecture bet is sound and will generalize with scale. The Skeptic thinks the whole pitch rests on the one task hardware can actually survive, and the hour-long autonomy is a slide, not a deployment. That gap is the decision. The Open-Source Advocate and the Compute Pragmatist agree on something darker for buyers: physical AI will not commoditize the way text models did, because the data and the real-time hardware both resist copying. If they are right, the winners here are a short list of full-stack companies, and the cheap-open-model dynamic that is squeezing your LLM bills does not arrive in robotics any time soon.

What it hinges on

One fact decides the story: does Dyna add a genuinely different task at a genuinely different customer, or does it keep selling variations of folding flat objects in kitchens? Napkins to vegetable cutting at the same chain is the same environment, the same cameras, the same lighting. That is adaptation, not generalization. A second customer, a different physical job, same base model, lightly fine-tuned, would prove the world-action thesis. Anything less proves they built a very good napkin machine. For an operator, the simpler read is this: watch whether "general-purpose physical agent" ships general-purpose work, or whether every deployment is a bespoke fine-tune that gets marketed as a platform.

Prediction: By the AI Engineer CODE conference in SF in November 2026, Dyna Robotics will publicly demo or announce only flat-object manipulation tasks (napkins, towels, laundry, basic cutting), not a structurally different physical task class like assembly, cooking with heat, or fragile-object handling, at a paying customer.

Confidence: Medium. The hand-overheating problem and the single-customer track record both point to narrow tasks staying narrow.

Why: The only scaled, revenue-attached deployments in the entire conversation are napkin-folding and towel-folding, with vegetable cutting named as the expansion, all flat or forgiving motions in controlled rooms. Ma himself names the two blockers: dexterous hands overheat and need ice packs during evaluation, and his vision-only model may need tactile sensing it does not yet have. Both constraints push hard toward tasks that tolerate imprecision and short grip cycles, which is exactly the flat-object category. The opposite outcome, a shipped demo of delicate or heat-based manipulation at a real customer, would require solving the hand-durability problem he openly admits is unsolved, which is not happening in weeks.

Revisit by 2026-12-01: We're right if Dyna's public demos and customer announcements through the November AI Engineer CODE event stay within flat-object folding and basic cutting. We're wrong if Dyna shows or announces a paying-customer deployment doing assembly, cooking with heat, or fragile-object handling.