9 min read
On September 28, Dyna Robotics published an uncut video of a single robot working a hotel laundry room for an hour without a person stepping in. It loads towels into the washer, moves the wet load to the dryer, folds what comes out, stacks it and puts the stacks on a shelf. The company says one cycle runs about 79 steps and contains 13 points where the robot has to decide what to do next. It calls the robot and its software together Dyna-2.1.[1] On October 13, the AI podcast Latent Space released an interview with Jason Ma, Dyna’s co-founder and chief scientist, covering the same system and what the robots are doing in paying customers’ buildings.[2]
The success rate of that hour is not in the post. There is no count of attempts, no count of completed runs, no count of human interventions. The company says as much itself: it argues that scoring a robot on single short tasks no longer tells you whether it can cover a shift, and proposes measuring how long a workflow can run unattended instead.[1]
Trained on a million hours of human video
The model inside it is called Dyna-2, which the company describes as a world-action model. It predicts what the scene in front of the camera will look like over the next few seconds, then derives arm and hand movements from that prediction, much as a chatbot predicts the next word to build a sentence.
Earlier robot models mostly took a camera feed plus a spoken instruction and produced motion directly. Dyna-2 inserts the prediction step in between. Ma compared it on the podcast to a language model working through a problem before it answers.[2]
The pre-training data is a million hours of first-person video of people working with their hands, mostly recorded with head-mounted cameras. The company puts that at roughly 170 years of continuous waking experience. A much smaller amount of data recorded on the company’s own robots is mixed in.[3] Language models could be trained on text scraped from the internet, and no comparable archive exists of hands picking things up and folding them, so human video stands in for it.
Dyna also published what happened as it scaled that data from 1,000 hours to 10,000, 100,000 and a million. The mean normalized score on 14 real robot tasks went 20%, 28%, 45%, 53%. Turning a key in a lockbox stayed at 0% through 100,000 hours and reached 90% at a million. Following a spoken instruction correctly rose from 35% to 96%.[3] All of it is the company’s own evaluation, and neither the data nor the code is public.
79 steps turn 95% into 1.7%

A single motion that works 95% of the time sounds like a competent robot. Multiply 0.95 by itself 79 times and you get 1.7%, or two complete runs out of a hundred attempts. Dyna put that calculation in its own post, along with an estimate that the same assumption implies about 190 human interventions per 24 hours.[1] Running the same arithmetic at 99% per step gives 45.2% for the full cycle, and it takes 99.9% per step to reach 92.4%.
Because the penalty compounds that fast, a long job is decided less by first-try accuracy than by how reliably the robot recovers after something goes wrong. That is why Dyna’s video spends time on the robot picking up a dropped towel and rebuilding a stack after someone knocks it over.[1]

Dyna-2.1 wraps that model in three layers. At the top, a model that looks at the screen and answers in words (a vision-language model) picks the next task and keeps a written note of how far along the workflow is. Deciding to go fold towels while the washer runs is this layer’s job. The middle layer is Dyna-2, generating motion roughly five times a second. At the bottom sits the controller that actually drives the arms and wheels, sending joint angles and wheel speeds 100 times a second; it was trained by trial and error in simulation.[1] The body, called Taku, is a human-sized upper torso with two arms, riding on four steerable wheels instead of legs. Ma cited energy use, stability and safety as the reasons for wheels.[2]
The napkins are already a business
Separately from the research video, these robots work in operating businesses. At several US locations of the dumpling chain Din Tai Fung, they fold thousands of napkins a day. Ma said the chain has since signed a larger contract and now wants the robots cutting vegetables and folding dumplings as well. At Best Western properties they fold towels and laundry. Tens of robots are deployed today, and the goal for next year is hundreds.[2]
Dyna also gave numbers on how fast a new site comes online. For stationary tasks, some new locations reached production in as little as three days, and the share of customer acceptance tests passed with no site-specific tuning was 87% for Dyna-2 against 46% for the previous model.[3] Lighting, tables and materials vary from site to site, and that variation is the hard part of the business. As Ma put it, “What’s really hard is not putting your robot in one location or doing a pilot.”[2]
About the field as a whole, he said “robotics is still very demo-oriented.”[2] The same standard applies to his own company’s hour-long video. An uncut hour carries much more information than a cut highlight reel, and it still leaves out how many runs were attempted and how many finished.
Physical Intelligence, a rival lab, is working the same problem
A November 2025 paper from the robotics company Physical Intelligence takes a different route to the same failure. The team runs robots in the real world, collects the failures along with corrections made by human experts steering the robot remotely, and folds all of it back into training. On their hardest tasks they report more than twice as much work finished per hour and roughly half the failure rate. The evaluation tasks were folding laundry in real homes, assembling boxes, and making espresso drinks on a commercial machine.[4]
Both companies answer the same way: alongside a bigger model, train it on what to do after it gets something wrong. Speeding up one step and finishing a whole job are different problems in software too, which is the same wall that always-on agents meant to work unattended in the cloud run into.
Factory robots are everywhere already, and they do a different job
The International Federation of Robotics reported in April that for 2024, South Korea had the world’s highest robot density at 1,220 industrial robots per 10,000 manufacturing employees, against a global average of 132 and 307 in the United States.[5] Those machines are bolted in place and repeat a fixed motion. The serving robots now common in restaurants carry food along a set path and never pick anything up.
Dyna is going after the work in between: repetitive tasks where the object lands in a different position every time, which is why they still needed hands. The company raised $120 million in September 2025, led by RoboStrategy, CRV and First Round Capital, with NVIDIA’s venture arm, the Amazon Industrial Innovation Fund, Samsung Next and LG Technology Ventures also taking part.[6]
What to check in the next robot video
More of these videos are coming, so a short checklist helps.
- Does the success rate come with a denominator. “99% success” means nothing without how many runs out of how many.
- How many times did a person step in, and how many hours did it run continuously. Deployment turns on covering a shift, not on one clean attempt.
- Does it work somewhere the robot has never been. Lighting and work surfaces differ between a demo floor and a working kitchen.
- Was the video cut, and was it sped up. A reel of successful moments carries much less information than an uncut run.
For workers, what a robot takes first is one repetitive task inside a job. Napkin folding and towel folding go first because they repeat hundreds of times a day and the finished product has one correct shape. Restaurant work is more likely to shift than to disappear, with less time spent on repetition and more on customers and exceptions. That read is based on the tens of robots deployed so far, and deserves another look when the number reaches hundreds.
If you are evaluating one of these for a business, note that Dyna has not published a price. The figures circulating in secondary coverage disagree with each other, and there is no official number.
A closing thought
An uncut hour is persuasive. Whether that hour was one run out of ten or three out of three changes what the business actually is. The company is right that scores on short tasks do not capture long work, and it still owes us numbers for whatever replaces them.
The firmer information right now is in the contracts. A paying customer expanding its deal and asking for vegetable cutting tells you something a success-rate table cannot. Watching which businesses start paying for which tasks is a better way to follow this field than counting how closely a robot resembles a person.
—
Sources
- Dyna Robotics, “Dyna-2.1: A Physical Agent for End-to-End Workflows” (2026-09-28)
- Dyna Robotics, “Dyna-2: A 1-Million-Hour Scaling Law for World-Action Models” (2026-08-15)
- Latent Space, “Dyna-2: General Purpose Physical Agents with Hour-Long Autonomy — Jason Ma, Dyna Robotics” (2026-10-13)
- Physical Intelligence, “π*0.6: a VLA That Learns From Experience” (arXiv:2511.14759)
- IFR, “Robot Density Surges in Europe, Asia, and Americas” (2026-04-08)
Notes
- About 79 steps and 13 decision points per hotel laundry cycle; three layers (whole-body controller at 100Hz, Dyna-2 policy at about 5Hz, vision-language orchestrator); an illustrative 1.7% cycle completion rate at 95% per-step reliability and an estimated 190 interventions per 24 hours. The company labels both figures illustrative rather than measured, and publishes no success rate, intervention count or trial count for the hour-long run. Dyna Robotics, Dyna-2.1 (2026-09-28) ↩
- Jason Ma interview. “We have already deployed tens of robots, but our goal next year is overall just deploying hundreds of robots.”, “our robots fold thousands of napkins per day”, “they not only want our robots to fold napkins. Now they want our robots cut.”, “What’s really hard is not putting your robot in one location or doing a pilot.”, “robotics is still very demo-oriented.” Latent Space (2026-10-13) ↩
- Pre-training on one million hours of egocentric human video (“roughly 170 years of continuous waking experience”); mean normalized score on 14 robot tasks of 20%, 28%, 45% and 53% at 1k, 10k, 100k and 1M hours; lockbox key turning 0% at 100k and 90% at 1M; instruction following 35% to 67% to 96%; zero-shot customer acceptance tests passed 46% for Dyna-1 and 87% for Dyna-2; new stationary sites in production in as little as three days. Company-run evaluations; data and code not released. Dyna Robotics, Dyna-2 (2026-08-15) ↩
- Physical Intelligence and 56 co-authors, RECAP (RL with experience and corrections via advantage-conditioned policies). “more than doubles task throughput and roughly halves the task failure rate”; evaluation on laundry folding in real homes, box assembly and espresso drinks. arXiv:2511.14759 (2025-11-18) ↩
- 2024 manufacturing robot density: South Korea 1,220 per 10,000 employees (highest in the world, growing 7% a year on average since 2019), global average 132, United States 307. IFR press release (2026-04-08) ↩
- $120 million Series A announced September 2025, led by RoboStrategy, CRV and First Round Capital, with Salesforce Ventures, NVIDIA’s NVentures, the Amazon Industrial Innovation Fund, Samsung Next and LG Technology Ventures participating. A $23.5 million seed round preceded it in March 2025. The Robot Report (2025-09-16) ↩
Source: https://www.dyna.co/dyna-2.1