Dyna's Lindon Gao says one working workflow equals all of them, and has moved spending toward pretraining

Lindon Gao, founder and CEO of Dyna Robotics, gave Zhang Xiaojun a 171-minute interview on episode 154 of her podcast 商业访谈录, released this week, and said that getting one scenario fully working is the same as getting all scenarios working. The bet behind that line is a budget shift: after a napkin product that met its spec but needed a human babysitter, Dyna moved from an even split of pretraining and post-training to spending much more on pretraining, a change Gao dated to this year. Dyna's own post gives the check, saying a laundry cycle chains about 79 steps and that at 95% per step it "almost never finishes without help."

The show notes contrast Dyna's route, which treats a workflow as the smallest unit and drives model pretraining from real tasks, with the pure brain research route of Physical Intelligence, and Gao put the gap between doing one workflow and doing 100 at "just the data for 100 workflows."

The napkin robot met its spec and still needed a babysitter

Dyna shipped Dyna-1 in June 2025 and deployed it to fold napkins from June to September. Gao said the robot handled four of the five folds and left the hardest, the last fold, to restaurant staff; the customer, a very large restaurant account that needed 1,300 napkins a day, complained once it saw how much staff time that last fold took. Dyna-1 folded 850 napkins in 24 hours at the start, about 65% of that daily need, and speed work brought it to target. Grabbing two napkins at once cost about two minutes to untangle, and Gao said a 10% double-grab rate would put the daily target out of reach.

The napkin product had more than 10 customers, and Gao said he decided in 2025 not to scale it because the robot finished the task without finishing the whole workflow. He called "the robot needs a babysitter" a very big product problem.

Gao put the reasoning in one line: "If you scale an incomplete product, it becomes a burden later." Asked about research labs that say the fundamental problem is the robot brain, he asked back, "If the brain falls short, where?" and then, "But in the napkin-folding scenario, was it a brain problem?" Within the task, he said, intelligence was not the constraint; the scenario failed because no closed loop formed across the whole workflow.

Pretraining grows, and Gao reverses his own rule on data quality

In 2025, Gao said, foundation models were not good enough to deploy: slow, messy, able to do one thing successfully once, so he set the team a 99% success rate as a research target. Dyna-1 used what Gao called the highest-quality data, and he told the team to go all in on it. He says he selectively heard what he wanted from language-model and video-generation researchers who reported that removing some data improved pretraining.

Workflows changed the rule. Once Dyna treated one workflow as the unit, Gao said the priority became whether the model sees the most diverse and broadest human experience, and that heavy quality optimization before reaching internet-scale data is "not necessarily first principles." He said that around the end of 2025 Dyna went through a period of strategic confusion, one group favoring deployment and another research, and he leaned toward research. This year the split moved, with pretraining now taking much more of the budget than post-training, and he said most of a model's robustness comes from pretraining.

A 79-step cycle needs near-99.9% per step to finish unattended

Treating those 79 steps as independent and allowing no recovery, a cycle completes about 1.7% of the time at 95%; at 99% the figure is about 45%, and at 99.9% about 92%. Dyna's 79-step count is an estimate and real systems recover from some failures. On this arithmetic, one workflow covers one unattended shift only when per-step reliability sits near 99.9%.

Dyna introduced Taku on Sept 29, a semi-humanoid with a human-shaped upper body, a folding lower body, four steerable wheels and two 7-degree-of-freedom arms on low-ratio planetary actuators, and showed an uncut, autonomous one-hour laundry-room workflow. The stack runs in three layers: a whole-body controller trained with reinforcement learning in simulation, running at 100 Hz; the Dyna-2 policy, a world-action model that produces whole-body target trajectories; and a vision-language-model orchestrator that tracks the workflow, keeps a text memory and picks the next step. Dyna counts thirteen decision points in a laundry cycle.

The pretraining bet has its own numbers. Dyna-2, published in August 2026, is pre-trained on more than one million hours of egocentric human video, roughly 170 years, and its mean normalized score across 14 tasks rose 20% to 28% to 45% to 53% as that human data went from 1,000 to 10,000 to 100,000 to 1,000,000 hours: a 33-point gain across a thousand-fold increase. At customer sites neither model had seen, Dyna reports a production pass rate of 46% for Dyna-1 and 87% for Dyna-2, 41 points apart.

Each workflow still needs its own recovery work

Gao's own account limits how far the data argument reaches. Placing napkins neatly at a given position in the bin is a problem pretraining can solve through language, he said, and Dyna traced its VLA's weak language-following to pretraining; grabbing two napkins at once is a problem pretraining cannot solve, so a workflow still needs post-training work on its specific failure.

The 87% figure is real customer-site evidence, but it comes from stationary tasks, where Dyna also says a new site reaches its production bar in as little as three days; the full workflow has no comparable number yet.

Dyna's two September documents disagree on how far deployment has gone. The technical post calls real-world customer sites "our next milestone," while the same-day press release says Dyna 2.1 is being deployed in hotels, laundromats and restaurants. Gao said the workflow has commercial customers but has "not fully deployed yet," naming hotels. He said each hotel laundry room needs 2 to 5 people depending on the hotel, that the approach could reach 100,000 US hotels, and put that market at "several billion," said robots are "one to two years" from scalable product-market fit, and said Dyna already beats his value metric: a robot working 24 hours matching a US worker's 8-hour day.

If the 87% holds on the full laundry workflow at customer sites, the pretraining shift will look like the thing that bought it.

In the 2025 napkin build, the robot sometimes sat idle for half an hour straight, Gao said, because staff had not put a new stack in place. In September, Taku ran a laundry room on its own for an uncut hour.