Open dataset for cloth‑folding robots

Cloth is one of the hardest things to teach a robot because a T-shirt does not keep one shape. A mug stays a mug, but a shirt turns into a different puzzle every time you pick it up. (huggingface.co) That is why most robot demos avoid laundry and stick to rigid objects like boxes, cups, or parts on a factory tray. Hugging Face’s LeRobot team picked the messy version on purpose and released the full recipe instead of just a highlight video. (huggingface.co) The setup uses two robot arms, which is called bimanual control. It is the robotics version of trying to fold a shirt with both of your hands instead of pinning one sleeve down and hoping for the best. (huggingface.co) The team says it used 8 bimanual robot setups, collected about 131 hours of demonstrations, and spent more than 5,000 graphics processing unit hours on training runs. That is a lot of trial-and-error to produce one task that looks simple in a bedroom and turns nasty in a lab. (huggingface.co) Those demonstrations are the part where a person shows the robot what “good folding” looks like. In robot learning, that teaching data plays the same role that labeled photos played for early image-recognition systems. (huggingface.co) LeRobot says the finished system reaches a 90 percent success rate on folding a random T-shirt. “Random” matters here because cloth robots often look good only when the shirt starts in the same pose every time. (huggingface.co) The release is bigger than one model checkpoint because it includes code, data, hardware details, teleoperation choices, training recipes, and evaluation notes. That means another lab can copy the pipeline, change one piece, and learn which part actually improved the result. (huggingface.co) A lot of robotics work still arrives as a polished video with missing ingredients. The LeRobot post says the goal here is to show every step from data collection to deployment so people can build on it instead of guessing how the demo was made. (huggingface.co) The timing also fits a larger push inside LeRobot to make robot data easier to store and stream. Its newer dataset format packs multiple episodes into single files and supports streaming large robot datasets without downloading everything first. (huggingface.co) That plumbing matters because robot datasets are not just folders of images. They combine video, arm positions, gripper actions, and timing data, and LeRobot’s format is built to keep those signals lined up frame by frame. (huggingface.co) LeRobot’s March 9, 2026 release also added support for the OpenArm robot and the OpenArm Mini teleoperator used in this folding project. So this cloth-folding drop lands as part of a broader attempt to turn robot training into something closer to open-source software than a private lab trick. (huggingface.co) The immediate result is not a home robot that will clear your laundry basket next week. The more concrete change is that students, startups, and research labs now have a public bimanual cloth-folding benchmark with working code, real robot data, and a reproducible baseline to beat. (huggingface.co)