How to build a laundry-folding robot

See how it goes together

The cheap setup is two SO-101 follower arms facing each other over the towel, a small camera on each wrist and one USB camera looking down from above. The exploded view names every part.

Pause

12345678910

What works and what doesn't

How the parts fit together

You teach the robot by doing the fold yourself through a pair of leader arms. The follower arms copy you while cameras record. A GPU turns those recordings into a policy, and the policy then drives the follower arms on its own. The software for all of it is LeRobot, Hugging Face's open-source robot learning library.

Parts list for the cheap setup

USD, as listed on each source page on Oct 23, 2026. Shipping and tax not included.

Step by step

1. 1. Learn the loop before you buy anything

   Install LeRobot, read the free Robot Learning Tutorial, and open a towel-folding dataset in the visualizer to see what a demo looks like: camera frames plus joint positions, usually 30 times a second.

   pip install lerobot

   LeRobot on GitHub, Robot Learning Tutorial (Hugging Face), LeRobot dataset visualizer

2. 2. Build and calibrate the arms

   Print or buy the parts, set each servo's ID, then calibrate every arm. The SO-101 guide walks through both leader and follower.

   lerobot-find-port
   lerobot-setup-motors ...
   lerobot-calibrate ...

   LeRobot SO-101 guide, SO-ARM100/101 repo (TheRobotStudio)

3. 3. Teleoperate both arms

   Move the leader arms and the followers copy you. LeRobot has a two-arm SO-101 setup built in (bi_so_follower with bi_so_leader). Practice until the folds look clean; early demos are worse than later ones.

   lerobot-teleoperate --robot.type=bi_so_follower --teleop.type=bi_so_leader ...

   LeRobot: imitation learning on real robots, LeLab, a web UI for LeRobot

4. 4. Record 50 or more demos

   Fix the cameras, use the same towel and table, and fold the same way each time. LeRobot suggests at least 50 episodes; the Hugging Face folding team suggests 30 to 50 clean demos per item and background.

   lerobot-record ... --dataset.single_task="Fold the towel in half" --dataset.num_episodes=50

   LeRobot: imitation learning on real robots, Unfolding Robotics (Hugging Face)

5. 5. Train a policy

   Start with ACT. It is small and trains on a laptop GPU. On a CPU-only machine, don't train locally: use the Colab notebook or rent a GPU.

   lerobot-train --dataset.repo_id=you/fold_towel --policy.type=act ...

   LeRobot compute and hardware guide, LeRobot Colab notebooks, ACT paper (arXiv 2304.13705)

6. 6. Let the policy drive

   lerobot-rollout loads your trained policy, reads the cameras and joints, and sends joint targets to the followers. Film every run so you can see where it fails, then record more demos of exactly those moments.

   lerobot-rollout --strategy.type=base --policy.path=you/my_policy ...

   LeRobot: imitation learning on real robots

7. 7. Level up: bigger models, smoother motion

   Try SmolVLA or π0.5 once ACT works. Real-time chunking (RTC) overlaps thinking with moving, which helps slow models. Relative actions and a reward model (SARM) were big wins in the Hugging Face t-shirt project.

   SmolVLA paper (arXiv 2506.01844), LeRobot real-time chunking, Unfolding Robotics (Hugging Face), openpi (Physical Intelligence)

The commands are shortened. The LeRobot imitation learning guide has the full flags for ports, cameras and the Hub. Prefer a browser UI? Try LeLab.

Folding demos and open data

• Unfolding Robotics: t-shirt folding ↗

  Hugging Face LeRobot team, 2026

  Bimanual OpenArm with OpenArm Mini leaders, three cameras, π0/π0.5 fine-tunes. 90% success over 20 rollouts: 100% on laid-out shirts (40.8 s average), 80% on messy shirts (95.9 s). About 131 hours of demos and over 5,000 GPU hours.

  Open: model, both datasets, hardware files, recipe. Model, Full dataset, High-quality dataset

• π0-ALOHA-towel checkpoint ↗

  Physical Intelligence, 2025

  An open checkpoint in openpi that "can fold diverse towels 0-shot on ALOHA robot platforms." Inference needs a GPU with more than 8 GB.

  Open: code and weights. The repo warns it may or may not work on your setup.

• π0 laundry folding ↗

  Physical Intelligence, Oct 2024

  A fine-tuned π0 takes crumpled clothes from a bin, folds them and stacks them, on a fixed two-arm robot and on a mobile robot that unloads a dryer first.

  Closed robot, paper public. The base π0 weights were opened later. π0 paper (arXiv 2410.24164)

• Helix folds towels ↗

  Figure, Aug 2025

  A Figure 02 humanoid with multi-fingered hands picks towels from a pile and folds them, end to end from camera to motors.

  Closed.

• Community towel-folding datasets on SO-101 arms ↗

  Hugging Face Hub users, 2025 to 2026

  Public LeRobot datasets of people teleoperating two SO-101 arms to fold towels: 100 episodes with 3 cameras, and 603 episodes with 4 cameras. There's also a 50-episode XLeRobot set. None publish a success rate we could check.

  Open datasets. 100-episode two-arm set, XLeRobot set

Folding machines that didn't make it

• Laundroid ↗ , Seven Dreamers, Bankrupt Apr 2019

  A $16,000 cabinet that took 5 to 10 minutes per shirt. It never shipped.

• FoldiMate ↗ , FoldiMate, Inc., Shut down 2021

  You clipped each garment in by hand. It never reached mass production, and deposits were refunded.

Open problems you can help with

• Getting from a pile to a flat item

  Folding a laid-out shirt hit 100% in the Hugging Face tests; starting from a messy shirt dropped to 80% and took more than twice as long.

  How to help: Record and share demos that start from a crumpled towel, not a flat one.

• Cloth doesn't simulate well

  The Hugging Face team says simulators can't reproduce how fabric crumples and slides, and world models can't yet predict it, so real teleop data is the only practical path today.

  How to help: Every clean, well-labeled real dataset you upload helps. Put the task name and camera layout in the card.

• Nobody reports cheap-arm success rates

  There are hundreds of public SO-101 towel episodes, but we found no published success rate for a policy trained on them.

  How to help: Run 20 filmed trials, count the successes, and post the numbers with the video.

• Gripping fabric

  The t-shirt project on OpenArm needed custom gripper jaws with a bigger contact patch to grip, pinch and slide fabric.

  How to help: Design a printable cloth-friendly jaw for the SO-101 and share the files.

• Big models need big GPUs

  π0-class fine-tunes need more than 22.5 GB of VRAM even with LoRA. ACT fits a laptop but is a much smaller model.

  How to help: Compare ACT and SmolVLA on the same towel dataset and share what each one gets right.

Papers behind the policies

ACT paper (arXiv 2304.13705), Diffusion Policy paper (arXiv 2303.04137), SmolVLA paper (arXiv 2506.01844), π0 paper (arXiv 2410.24164)