How to build a dish-loading robot

See how it goes together

XLeRobot is two SO-101 style arms on top of a wheeled IKEA RÅSKOG cart, with a camera head on a pan and tilt mount and a small camera on each wrist. A power station rides inside. Here it is coming apart and going back together next to a dish rack.

Pause

123456789101112

What works and what doesn't

Why XLeRobot

Dishes start on the table and end up in the sink or a rack, so the robot needs wheels and two hands. XLeRobot is the cheapest open design we found with both, it runs on the same LeRobot tools as the SO-101 arms, and there's already public table-clearing data recorded on it. LeKiwi has wheels but one arm. OpenArm has bigger arms but no base, and a two-arm kit starts around $6,500.

How the parts fit together

You teach the robot by doing the job yourself with a controller. The robot copies you while its cameras record. A GPU turns those recordings into a policy, and the policy then drives the arms and wheels on its own. The software is LeRobot, Hugging Face's open-source robot learning library, plus the XLeRobot plugins.

Parts list for XLeRobot

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

Step by step

1. 1. Watch real dish data before you buy

   Install LeRobot and open the public XLeRobot clean-table dataset in the visualizer. It has 50 episodes of an XLeRobot moving a cup and spoon off a plate, seen from the head and both wrists. That's what you'll be recording.

   pip install lerobot
   lerobot-visualize-dataset zonglin11/xlerobot_clean_table

   LeRobot on GitHub, xlerobot_clean_table dataset, LeRobot dataset visualizer

2. 2. Build the arms, the base and the head

   Follow the assembly guide and video. Set every servo's ID and calibrate the arms with the SO-101 guide, then the base and head motors with the XLeRobot guide. The kit ships the arms already built.

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

   XLeRobot assembly guide, Assembly video, LeRobot SO-101 guide

3. 3. Install the XLeRobot plugins

   XLeRobot ships as LeRobot plugins. After this, --robot.type=xlerobot works in the normal LeRobot commands. Pick the plugin that matches your base.

   git clone https://github.com/Vector-Wangel/XLeRobot.git
   cd XLeRobot
   pip install -e software/plugins/xlerobot_model
   pip install -e software/plugins/lerobot_robot_xlerobot   # or _2wheels
   pip install -e software/plugins/lerobot_teleoperator_xlerobot_vr

   XLeRobot install guide, XLeRobot repo

4. 4. Drive it by hand and carry a plate

   Start with the keyboard script, then try an Xbox controller or Joy-Cons. Drive to the table, pick up a plastic plate by the rim, drive to the counter and set it in the rack. If you can't do it by hand, a policy won't learn it either.

   python software/examples/4_xlerobot_teleop_keyboard.py
   python software/examples/5_xlerobot_teleop_xbox.py
   python software/examples/7_xlerobot_teleop_joycon.py

   XLeRobot teleop guide

5. 5. Record 50 demos of one small job

   Pick one job, like "cup from the mat into the rack". Park the base in the same spot, keep the lighting and the rack fixed, and record 50 clean episodes. The XLeRobot guide records with a leader arm or a Quest 3, and lets you delete a bad episode from the controller.

   lerobot-record --robot.type=xlerobot --teleop.type=xlerobot_vr \
     --dataset.single_task="Put the cup in the rack" --dataset.num_episodes=50 ...

   XLeRobot ACT guide, LeRobot: imitation learning on real robots

6. 6. Train a policy

   Start with ACT. It's small and trains on a laptop GPU. SmolVLA starts from a pretrained base and wants a bigger GPU; the XLeRobot guide has the command.

   lerobot-train --dataset.repo_id=you/cup_to_rack --policy.type=act ...
   lerobot-train --policy.path=lerobot/smolvla_base --dataset.repo_id=you/cup_to_rack ...

   LeRobot compute and hardware guide, XLeRobot SmolVLA guide, ACT paper (arXiv 2304.13705)

7. 7. Let it run, and film every try

   ACT can run on the robot's own computer. For SmolVLA or π0.5, run a policy server on a GPU machine and let the robot connect over your network. Count successes over 20 tries, then record more demos of the moments it fails.

   lerobot-rollout --strategy.type=base --policy.path=you/my_policy ...
   python -m lerobot.async_inference.policy_server --host=0.0.0.0 --port=8080

   XLeRobot ACT guide, LeRobot async inference, LeRobot: imitation learning on real robots

8. 8. Level up: chain jobs together

   Once a few single jobs work, string them together: drive, clear the cup, clear the plate, drive back. A community π0.5 port for two-arm XLeRobot exists, and the docs show an LLM agent calling trained policies as tools.

   XLeRobot π0.5 guide, openpi (Physical Intelligence), XLeRobot docs

The commands are shortened. The XLeRobot ACT guide and the LeRobot imitation learning guide have the full flags for ports, cameras and the Hub. Want to practice first? XLeRobot also has a simulation.

Dish robot demos and data

• XLeRobot clean-table dataset and models ↗

  Hugging Face Hub user zonglin11, Jun 2026

  50 episodes (121,451 frames at 30 fps) of a two-wheel XLeRobot moving a cup and spoon off a plate on a table mat, with head and wrist cameras. SmolVLA and X-VLA models trained on it are public. No success rate is published.

  Open: dataset (Apache 2.0) and models. SmolVLA model, X-VLA model

• XLeRobot 0.3.0 home chores ↗

  XLeRobot project, Aug 2025

  Ten chores including wiping a table, getting drinks from the fridge, filling a cup with ice and picking from drawers. Every clip is teleoperated with Joy-Cons and sped up. None of them load dishes.

  Open hardware and code, Apache 2.0. XLeRobot demos page

• Mobile ALOHA: Rinse Pan ↗

  Stanford, Jan 2024

  A $32,000 open-hardware mobile robot picks up a dirty pan, rinses it under the tap and puts it on the drying rack. 80% success from 50 demos when co-trained with older ALOHA data, and 0% without it. The paper lists loading and unloading a dishwasher among example chores the hardware can do, but it didn't train a policy for that.

  Open: hardware, code and data. Mobile ALOHA paper (arXiv 2401.02117)

• π0.5 puts dishes in the sink in new homes ↗

  Physical Intelligence, Apr 2025

  A mobile robot clears dishes into the sink in homes that weren't in the training data. The team says it often makes mistakes and doesn't always succeed on the first try.

  Closed robot. The π0.5 base weights are open in openpi. π0.5 paper (arXiv 2504.16054), openpi (Physical Intelligence)

• Helix loads the dishwasher ↗

  Figure, Sep 2025

  A Figure 02 humanoid singulates stacked plates, hands a glass from one hand to the other and loads a real dishwasher, recovering from misgrasps. Figure notes the racks leave centimeter tolerance.

  Closed.

• Public dishwasher datasets ↗

  Unitree and RoboCasa users, 2026

  Unitree published 200 episodes of a G1 humanoid putting a plate into a dishwasher. There are also 501 simulated LoadDishwasher episodes from RoboCasa in LeRobot format. Neither is on XLeRobot, and neither reports a success rate.

  Open datasets. RoboCasa LoadDishwasher dataset (sim)

Dish robots that didn't make it home

• Bot Handy ↗ , Samsung, Shown at CES 2021

  A one-arm home robot shown loading a dishwasher. Samsung said it was in development, not for sale, and shown in test settings. We found no sign it ever shipped.

• Dishcraft ↗ , Dishcraft Robotics, No longer operating

  Dishcraft built central washing hubs for businesses that use reusable dishware, not home robots. It raised more than $48M, and Unreasonable Group's venture page now lists it as no longer operating.

Also see The Verge on Bot Handy.

Open problems you can help with

• No one has shown XLeRobot loading dishes

  The closest public work is the clean-table dataset, which moves a cup off a plate. We found no XLeRobot clip of a plate going into a rack or a dishwasher, teleoperated or not.

  How to help: Record a dish-to-rack dataset on XLeRobot, put the camera layout in the card, and post a video.

• Reach into a real dishwasher

  The arms sit on top of a waist-high cart. A dishwasher's bottom rack is near the floor. We couldn't confirm the arms reach it, which is why this plan aims at a countertop rack or the sink. Even the $32,000 Mobile ALOHA lists its lowest reach as 65 cm.

  How to help: Measure the reach of a stock XLeRobot and share it, or design a lower arm mount.

• Wet, slippery, breakable things

  Figure calls out slippery and fragile items as a core difficulty. The SO-101 gripper is two printed jaws, and nobody publishes how it does on wet ceramic.

  How to help: Test grip pads or a printed plate-friendly jaw and share the files and numbers.

• Long jobs need many steps to go right

  Clearing a table is drive, grasp, carry, place, repeat. Mobile ALOHA hit 80% on the single pan job only with extra co-training data, and dropped to 0% without it.

  How to help: Report per-step success, not just whole-task success, so others can see which step breaks.

• Nobody reports cheap-robot success rates

  Public XLeRobot and SO-101 datasets keep growing, but we found no published success rate for a dish task on them.

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

Papers behind the policies

ACT paper (arXiv 2304.13705), SmolVLA paper (arXiv 2506.01844), π0.5 paper (arXiv 2504.16054), Mobile ALOHA paper (arXiv 2401.02117)