Dyna Robotics’ Taku runs a one-hour hotel laundry shift on Dyna-2.1

Dyna Robotics unveiled Taku, a wheeled semi-humanoid, and Dyna-2.1, its new control stack, on September 29, 2026, with a video of the robot running a hotel laundry workflow autonomously for one uncut hour — about 79 sequential steps of washing, drying, folding and stacking, by the company’s count. The length is the point: most manipulation demos show one task in a short clip, while the failures that stall deployments tend to happen between tasks, when a robot has to decide what to do next.

Dyna calls the result “the first physical agent to automate an end-to-end workflow”. That is the company’s own framing, and the run was a single demonstration on Dyna’s own terms; it did not name a hotel or customer.

What did Dyna actually show?

Taku keeps human proportions above the waist — a 21-DoF upper body with two seven-degree-of-freedom arms driven by low-ratio planetary actuators — and replaces legs with a folding lower body on four steerable wheels. Dyna says the robot is sized to match average human reach. The trade is deliberate: a wheeled base gives up stairs and rough ground in exchange for stability while the arms work, which suits back-of-house rooms with flat floors.

In the laundry run, the robot washed, dried, folded and shelved towels, and chose the order of work itself. Humanoids Daily noted that it also handled recovery, for example returning a dropped item to the dirty pile rather than folding it. Dyna says the next milestone is putting the new system into real customer sites.

How does the three-layer Dyna-2.1 stack work?

Dyna-2.1 splits control by speed. At the bottom, a whole-body controller trained with reinforcement learning in simulation runs at 100 Hz and keeps the body coordinated. In the middle, an improved Dyna-2 manipulation policy, which Dyna describes as a world-action model, runs at about 5 Hz. On top, a vision-language orchestrator decides which task comes next and keeps a text-based memory of where the workflow stands.

The layering resembles the fast-and-slow designs used elsewhere, including Figure’s Helix; our explainer on VLA models covers the general pattern. The detail worth noting is the text memory: a workflow state written in plain language can be inspected and corrected by an operator, which matters more in an hour-long shift than in a thirty-second clip. Dyna says the policy layer is pretrained on one million hours of human video mixed with robot data, using a shared motion format it calls Unified Robot Representation so that learning carries across bodies.

What is still missing before hotels buy this?

Dyna has not published success rates across repeated runs, throughput compared with a human laundry attendant, or the number of interventions a typical shift would need. One uncut hour shows the stack can hold together; it does not show how often it does. Those are the figures that will decide whether a hospitality operator signs a contract.

The company has deployment history to build on. When it raised a $120 million Series A in September 2025 — led by Robostrategy, CRV and First Round Capital, with Salesforce Ventures, NVentures, the Amazon Industrial Innovation Fund, Samsung Next and LG Technology Ventures participating — Dyna said its earlier DYNA-1 model was already running at customer sites. That round followed a $23.5 million seed in March 2025. We found no newer funding confirmed as of October 7, 2026. Taku is new hardware, though, and a new body resets much of the reliability work.