A robot doing a load of laundry has to do more than fold a towel. According to Jason Ma, co-founder and chief scientist of Dyna Robotics, it has to notice that the washer is still running, find something else useful to do in the meantime, and then move the towels into the dryer as soon as the wash finishes. If it waits, "that becomes a bottleneck in the whole workflow," he said.
Ma described that example on the Latent Space podcast, in an episode published on October 13, 2026. He used it to explain the company's argument about what robots need in order to work unsupervised for an hour instead of handling a 30-second task.
From short tasks to whole workflows
Dyna's core model is what it calls a world-action model. It is trained to predict how a scene will change, and it uses that prediction to produce the movements the robot actually makes. Ma said that kind of model works at a low level: it reasons about continuous motion. A real job also calls for a different kind of thinking.
"When we are doing decision making we think about steps of a recipe," Ma said. People think about what they have finished, what comes next, and where to go while something else is still running. He said a robot needs that high-level decision-making even for the laundry job.
In theory, Ma said, a single low-level model could learn all of that, but doing so would be "extremely inefficient." Dyna instead puts an agentic harness around the world-action model. A harness is a software layer that plans the job, keeps track of progress and calls on the robot's individual skills. Ma described this as a hierarchical system that lets the model compose different capabilities and skills to solve a much longer-horizon task.
Ma compared the change to what has happened with AI coding tools. Two or three years ago, he said, coding models were mainly good at autocomplete, and people had to keep prompting them to produce more code. Today agent systems carry a task from start to finish. In his view, robotics still mostly thinks about autonomy "at the very task level," meaning tasks of a minute or 30 seconds, while the real value is in workflows that run unsupervised for hours.
Asked whether the harness is custom for each use case or a product of its own, Ma said Dyna will eventually have its own SDK, a software kit for building on its system. For now, he said, the company wants a general harness layer for robotics foundation models, which today means its own model. He described general-purpose robotics as a systems problem more than a modeling problem: you need a good agent, a good harness, a good model and good hardware.
Ma also said context matters. A robot that decides only from what it sees at a given moment can get confused. Halfway through opening a microwave, he said, it is unclear from a single view whether the robot is opening the door or closing it.
What the company says the laundry system does
Dyna's Dyna-2.1 release, dated September 29, 2026, sets out this design for its robot, Taku. The robot has four steerable wheels, two arms and an articulated body that lets it reach into dryers and up to high or down to low shelves. The company describes three layers of control:
- a whole-body controller, trained in simulation, that runs 100 times a second;
- a Dyna-2 action policy that carries out the physical skills;
- an orchestrator that combines vision and language and keeps written notes on the state of the machines and the job's progress.
In Dyna's account, folding can pause when a dryer finishes, which matches Ma's point about not letting a machine sit idle. The cycle illustrated on the page has about 79 steps and 13 decision points, so the system has to be able to recover from mistakes along the way. The company presents an uncut, hour-long laundry demonstration and names deploying the new system at a customer site as its next milestone.
What still limits longer, more general work
The conversation then turned to harder jobs: setting up a filming studio by moving tables and lights, plugging things in, managing a data center, recharging another robot, or working through a five- to ten-minute sequence where many more things could go wrong.
Ma said length alone is no longer the main problem. The laundry demo takes about an hour, so the longest full job the robot can do "is already in the hour range." The obstacle to doing the same for any task, he said, is "coming back to data and hardware."
Hands. Setting up a room might need something more dexterous than a parallel-jaw gripper, the simple two-fingered pincer that opens and closes. Something closer to a human hand would help, but Ma said the best dexterous hands today "are still not durable enough. They overheat pretty quickly."
Slow testing. Limited hardware slows everything else, Ma said, because "to know that you're making progress on any task you eventually have to eval on the hardware." Tests on a physical robot cannot be run quickly and on demand, so research moves more slowly, and it becomes harder to work out which data to collect for very complex jobs like the studio setup.
Missing senses. Most robot training data today is visual, Ma said: video streams plus some record of hand or joint motion. People also hear, feel and smell things. "We subconsciously use all of those," he said. If that information is not captured, then "even when we scale to millions of hours you might just be missing critical dimensions of the task."
These gaps are what Ma sees "dragging on the timeline" for truly general robots. For workflows where high-fidelity data can already be collected, he said he does not see a fundamental bottleneck. His example was the laundry agent, which handles the whole workflow as a human would without needing more than a parallel-jaw gripper or any sensing that is not available today.