Towel-folding robot reported at 99.5% success on a laundry floor

Lukas Ziegler posted about @Gobanorobotics, a team out of Nantes and London running an autonomous towel-folding station on a real commercial laundry floor. The figures he relays are 99.5% task success, 68% of a human worker's throughput on an 18-hour shift, and roughly 5% throughput improvement per day driven by reinforcement learning (the robot improving from its own trial and error rather than fresh human teaching).

Lukas Ziegler

@lukas_m_ziegler

X

A towel bends, folds and slips the moment a gripper touches it, so a grasp that worked last cycle can miss or catch three layers on the next one.

Lukas Ziegler

@lukas_m_ziegler

X

The real unit of progress is the work done between human interventions

Those are the company's own numbers. Nothing in the source shows third-party verification, and no video evidence is described beyond the link Ziegler shares.

Why towels are hard

Cloth is the classic manipulation problem. As Ziegler puts it, a towel bends, folds and slips the moment a gripper touches it, so a grasp that worked on the last cycle can miss or catch three layers on the next one. There is no fixed shape to plan against, which is why folding demos usually run in a lab with one towel and a lot of resets.

The measurement is the interesting part

Ziegler's main praise is not the success rate, it is the bookkeeping. @Gobanorobotics tracks task success and autonomy separately, on the logic that a miss the robot recovers from by itself is a different event from a stoppage that pulls a human worker over to the machine. He argues the real unit of progress is the work done between human interventions, and says the throughput figure already includes the recovery time.

That matters because 68% of a human is a claim that gets much weaker if a person is standing nearby unjamming the gripper. Counting recovery inside the throughput number is the harder way to report it. Whether the number holds up is still unverified in the sources.

The learning loop

The company's reinforcement learning engine is called Toutatis. Per Ziegler, it starts from human demonstrations and a task-specific world model (a learned simulation of how the task behaves), then improves the policy from the robot's own experience on the floor. The claimed 5% daily throughput gain is the output of that loop.

This is the shape a lot of commercial robot learning is taking right now. Seed with demonstrations, deploy into one narrow setting, then let real shifts supply the data instead of a lab. A laundry is a good place to try it, since the task repeats endlessly, the failure cost is near zero, and the floor runs 18 hours a day.

Shown versus claimed, nobody outside the company has checked any of these three numbers yet.