Dyna-2.1: a semi-humanoid robot does uncut hotel laundry on camera

Dyna-2.1 is trying to do something robotics keeps promising and rarely delivering: take over a whole workflow, not just a neat demo. The system centers on Taku, a semi-humanoid robot with a human-shaped upper body, a folding lower body, a base on four steerable wheels, and two 7-degree-of-freedom arms. In the laundry setup, that body matters as much as the software. Washers, dryers, shelves and folding tables sit meters apart, and the robot has to reach deep into a drum, grab stray towels, turn with a load, crouch low, and reach overhead without a human stepping in.

That laundry workflow is the point of the paper. It is long, messy, and full of recoverable mistakes. Towels can snag, folds can slip, stacks can collapse, and one bad drop can send work back to the start. The authors break a cycle into thirteen decision points and say the machine has to keep watching the room while it folds, because a finished washer or dryer sitting idle is wasted capacity. They also argue that per-step reliability compounds brutally: if each step is only 95% reliable, a full cycle almost never finishes cleanly without help.

So Dyna-2.1 splits the problem into three layers. A whole-body controller runs at 100 Hz and turns target trajectories into joint targets and wheel speeds. A DYNA-2 policy turns the current step into those target trajectories. And a workflow orchestrator, built on a vision-language model, keeps track of what stage the laundry is in and chooses the next move. The system also uses a shared representation called URR, which describes wrist, elbow, chest and footprint poses so human recordings and robot motions can train the same stack.

The teachability angle is the other big claim here. The team says human video, motion capture and teleoperation can all be converted into URR, and that lets the robot learn from a lot more than just its own expensive hardware data. They say Dyna-2.1 learns new skills faster than earlier systems, with less demonstration data, and point to recent skills like server servicing and retrieving a drink. In their setup, a new stationary folding site can now reach its production bar in as little as three days, down from weeks to months of on-site engineering. That is the kind of number robotics people remember.

But the real message is simpler: the field is moving from isolated manipulation tricks toward full-stack robots that can survive the boring parts of work. The hard part is not making a robot pick up a towel once. It is making it keep going when the towel folds wrong, the machine finishes at the wrong time, and the workflow does not care about your demo.