How to build SO‑101 arm: the whole process
255 steps to build SO‑101 arm, from buying parts, printing and assembly to software, calibration and training, with the bill of materials, shop links and the official so…
About $229.88 (≈CN¥1,343) for a leader+follower pair; about $121.94 for a follower only. Printing and shipping excluded
Bill of materials · Parts For Two Arms (Follower and Leader Setup)
| Part | Qty | US | EU | CN | JP |
|---|---|---|---|---|---|
| STS3215 Servo 7.4V, 1/345 gear (C001) | 7 | USD 13.89 | EUR 12.2 | CNY 97.72 | JPY 2980 |
| STS3215 Servo 7.4V, 1/191 gear (C044) | 2 | USD 13.89 | EUR 12.2 | CNY 97.72 | JPY 2980 |
| STS3215 Servo 7.4V, 1/147 gear (C046) | 3 | USD 13.89 | EUR 12.2 | CNY 97.72 | JPY 2980 |
| Motor Control Board | 2 | USD 10.6 | EUR 11.4 | CNY 27 | JPY 980 |
| USB-C Cable 2 pcs | 1 | USD 7 | EUR 7 | CNY 47.8 | JPY 1498 |
| Power Supply | 2 | USD 10 | EUR 15.7 | CNY 22.31 | JPY 1550 |
| Table Clamp 4pcs | 1 | USD 9 | EUR 9.7 | CNY 20.8 | JPY 2200 |
| Screwdriver Set | 1 | USD 6 | EUR 9 | CNY 14.9 | JPY 500 |
Bill of materials · Parts for One Follower Arm
| Part | Qty | US | EU | CN | JP |
|---|---|---|---|---|---|
| STS3215 Servo 7.4V, 1/345 gear (C001) | 6 | USD 13.89 | EUR 12.2 | CNY 97.72 | JPY 2980 |
| Motor Control Board | 1 | USD 10.6 | EUR 11.4 | CNY 27 | JPY 980 |
| USB-C Cable 2 pcs | 1 | USD 7 | EUR 7 | CNY 23.9 | JPY 1498 |
| Power Supply | 1 | USD 10 | EUR 15.7 | CNY 22.31 | JPY 1550 |
| Table Clamp 2pcs | 1 | USD 5 | EUR 8 | CNY 7.8 | JPY 2200 |
| Screwdriver Set | 1 | USD 6 | EUR 9 | CNY 14.9 | JPY 500 |
Buy parts
- Source the robot components
Source the components needed to build the SO‑101 yourself from the Bill of Materials.
- Buy a kit or assembled arms
Buy assembled arms or a parts kit from one of the vendors linked in the document.
- Buy parts for two-arm teleoperation
When creating the classic teleoperation setup for use with the `LeRobot` library, buy the parts from the Parts for Two Arms section.
- Buy the leader-arm servos
The Alibaba bundle contains all six STS3215 servos needed for the SO-101 leader arm: 3 × 1/147 gear (C046), 2 × 1/191 gear (C044), and 1 × 1/345 gear (C001).
- Buy parts for one follower arm
Buy the off-the-shelf parts needed for one follower arm from this section.
- Choose the follower motor voltage
The 7.4V STS3215 follower motors were found to be sufficient; for more power, choose the 12V version. The 12V version also requires a 12V 5A+ power supply instead of a 5V one.
- Prepare Phillips screwdrivers
The exact screwdriver set is not required, but Phillips-head screwdrivers in sizes #0 and #1 are highly recommended for installing and removing screws.
- Build the XLeRobot dual-arm mobile robot
The XLeRobot has 2x SO101 arms, 1x Lekiwi base, 1x 300Wh Anker Battery, 2x wrist RGB cameras and 1x head depth camera (with a 2-dof neck), total cost 660$.
- Add AnySkin tactile sensing
Add touch sensing to your gripper by buying the AnySkin sensor from WOWROBO.
- Source the SO-101 parts
Follow the linked SO-ARM100 README to obtain the bill of materials and source the parts. The README provides links for sourcing the parts.
- Select a USB Camera Module
Prepare 1 USB Camera Module; the linked model is recommended, while any 32 × 32 mm module rated at least 720p/30 fps will likely work.
- Prepare M2 Screws
Prepare 4 M2 screws, which are the smaller screws supplied with the Feetech servos.
- Prepare M3 × 8 mm Screws
Prepare 2 M3 × 8 mm screws.
- Prepare M3 Hex Nuts
Prepare 2 M3 hex nuts.
- Get a 32 mm UVC camera module
Obtain one USB Camera Module. The linked model is recommended; a compatible 32 mm × 32 mm module with at least 720p/30 fps will likely work.
- Prepare four M2 screws
Prepare four M2 Screws, which are the smaller screws supplied with the Feetech servos.
- Prepare the camera and screws
Prepare one 32mm x 32mm camera module and 16 M2 screws that came with the Feetech servos.
Print and make
- Print or order the parts
3D print the parts or order the 3D-printed parts using the Printing the Parts instructions.
- Choose a 3D Printer
Use an FDM printer for the follower and leader arm parts. Suggested examples include the Prusa MINI+, UP Plus 2, Creality Ender 3, and Bambu Lab A/P/X-series.
- Use PLA+ Filament
Select PLA+ as the printing material.
- Set Nozzle and Layer Height
Use a 0.4mm nozzle with a 0.2mm layer height, or a 0.6mm nozzle with a 0.4mm layer height.
- Set 15% Infill Density
Set the infill density to 15%.
- Clean the Print Bed
Clean the print bed so it is free of dust or grease. If water or another liquid is used, dry the bed.
- Apply Bed Adhesive if Recommended
If recommended by the printer, apply a thin, even layer of standard glue stick across the print area, avoiding clumps or uneven application.
- Load Printer Filament
Load the printer filament using the printer-specific instructions.
- Match Suggested Printer Settings
Adjust the printer settings to match the suggested values above as closely as possible.
- Configure 45-Degree Supports
Enable supports everywhere, but ignore slopes greater than 45 degrees to the horizontal.
- Exclude Supports from Horizontal Screw Holes
Ensure that screw holes with horizontal axes contain no supports.
- Remove printed parts from the bed
Once printing is done, use a putty knife to scrape the parts off the print bed, then remove any support material from the parts.
- Use a printing service
If you do not own a 3D printer, follow the linked printing services page.
- Extend with optional hardware
These add-ons extend your SO-100/SO-101 robot.
- Print the mount helper jig
Print the handy mount jig for easier alignment during assembly; see its README.
- Print overhead camera mounts
For bird's-eye views in single or bi-manual setups, print either the Webcam version or the 32x32 Cam Module version using the linked instructions.
- Print the base mounts
Download the Raised Leader Base STL and follow the instructions for the 4040 Aluminum Profile Mount.
- Print the compliant gripper
Download the compliant gripper parts and print them with a flexible filament like TPU 95A; see the README for details.
- Print the new compliant gripper
The new gripper uses TPU 95A for the finger and PLA for the base, needs no support for the TPU finger, and requires 2 additional M3 screws plus optional 3M gripper tape for higher friction.
- Upload the SO-101 print files
Upload `STL/SO101/Leader/Ender_Leader_SO101.stl` for the Leader and `STL/SO101/Follower/Ender_Follower_SO101.stl` for the Follower on PCBWay.
- Set the print order options
Set the quantity to 1 for both, the design units to mm, and Custom material to `PLA+`. Set Product Desc to DIY Entertainment → Robot components.
- Specify FDM printing and submit
Enter `FDM, 20% infill` under Other special requests and optionally specify a color. Verify everything and click `Submit`.
- Wait for verification and review the quote
After the parts are verified, PCBWay provides a final quotation including shipping costs. You can contact PCBWay through its online portal.
- Prepare the 3D-printed parts
Follow the linked SO-ARM100 README for the instructions to 3D print the parts. Use its advice if this is your first time printing or you do not own a 3D printer.
- Download the base mount model
Download the STL file for the 4040 base mount for 3D printing.
- Download the STEP model
Download the STEP file for the 4040 base mount.
- Print in the suggested orientation
For optimal strength, print the mount in the orientation shown in the image.
- Enable print supports
Supports are required when printing this part.
- Print the compliant gripper
Print with a Bambu Lab printer using TPU filament with Shore hardness of 95A; the printer settings included 20% and support generation.
- Remove the printed supports
After printing concludes, remove the supports manually with flush cutters.
- Print both gripper components
Print Compliant_Moving_Jaw_SO101.stl and Compliant_Wrist_Roll_Follower_SO101.stl using TPU 95A.
- Print the Camera Adapter
Print 1 Hex-Nut Recess Camera Adapter in the orientation shown in the STL using tree supports. A 40% infill is recommended for sturdiness and to avoid wobbling.
- Print the Hex-Nut Camera Adapter
3D print the Hex-Nut Recess Camera Adapter.
- Print the mounts and arm base
Print one Arm Base per follower arm, plus one Camera Mount Bottom, one Camera Mount Middle, and one Camera Mount Top.
- Print the robot-specific wrist replacement
Print one Wrist Roll Replacement, also described as the 3D-printed Jaw Replacement Part. Use the corresponding STL file for SO-100 or SO-101.
Assemble
- Select the motors for each joint
Use the motor table to identify which motor is needed for each joint. The follower uses six STS3215 motors with 1/345 gearing, while the leader uses three differently geared motors to sustain its own weight and move with little force.
- Connect the gripper and wrist roll motors
Disconnect the controller board’s 3-pin cable while leaving it connected to the gripper motor. Connect another 3-pin cable between the wrist roll motor and controller board, ensuring this is the only motor connected and that the motor is not connected to another motor.
- Chain the motor cables
Plug the 3-pin cable from each motor into the next motor. Connect the shoulder pan motor, with id=1, to the controller board and attach the board to the arm base.
- Remove support material
Remove all support material from the 3D-printed parts. Use a small screwdriver underneath the support material to lift it away.
- Install a cable in each placed motor
After placing the motors, install one 3-pin cable in each motor before continuing the assembly.
- Install the first motor horns
Install both motor horns, securing the top horn with a M3x6mm screw; the bottom horn needs no screws.
- Place the first motor
Place the first motor into the base.
- Fasten the first motor
Fasten the motor with four M2x6mm screws: two from the top and two from the bottom.
- Attach the first motor holder
Slide over the first motor holder and fasten it with two M2x6mm screws, one on each side.
- Attach the shoulder part
Attach the shoulder part.
- Tighten the shoulder part
Tighten the shoulder part with four M3x6mm screws on top and four on the bottom.
- Add the shoulder motor holder
Add the shoulder motor holder.
- Install the second motor horns
Install both motor horns, securing the top horn with a M3x6mm screw; the bottom horn needs no screws.
- Slide in the second motor
Slide the second motor in from the top.
- Fasten the second motor
Fasten the second motor with four M2x6mm screws.
- Attach the upper arm
Attach the upper arm with four M3x6mm screws on each side.
- Install motor 3 horns
Install both motor horns, securing the top horn with a M3x6mm screw; the bottom horn needs no screws.
- Insert and fasten motor 3
Insert motor 3 and fasten it using four M2x6mm screws.
- Connect the forearm
Connect the forearm to motor 3 using four M3x6mm screws on each side.
- Install motor 4 horns
Install both motor horns, securing the top horn with a M3x6mm screw; the bottom horn needs no screws.
- Slide over motor holder 4
Slide over motor holder 4.
- Slide in motor 4
Slide in motor 4.
- Fasten motor 4
Fasten motor 4 with four M2x6mm screws.
- Install motor 5 in wrist holder
Insert motor 5 into the wrist holder and secure it with two M2x6mm front screws.
- Install the wrist motor horn
Install only one motor horn on the wrist motor and secure it with a M3x6mm horn screw.
- Secure the wrist to motor 4
Secure the wrist to motor 4 using four M3x6mm screws on both sides.
- Attach the Gripper
Attach the gripper to motor 5 and to the motor horn on the wrist using 4 M3x6mm screws.
- Secure the Gripper Motor
Insert the gripper motor and secure each side with 2 M2x6mm screws.
- Install the Gripper Motor Horns
Install both motor horns on the gripper motor. Secure the top horn with a M3x6mm screw; the bottom horn requires no screws.
- Install the Gripper Claw
Install the gripper claw and secure both sides with 4 M3x6mm screws.
- Mount the Leader Holder
Mount the leader holder onto the wrist and secure it with 4 M3x6mm screws.
- Attach the Leader Handle
Attach the handle to the leader holder using 1 M2x6mm screw.
- Install the Leader Gripper Motor
Insert the gripper motor, secure each side with 2 M2x6mm screws, and attach a motor horn with a M3x6mm horn screw.
- Attach the Follower Trigger
Attach the follower trigger with 4 M3x6mm screws.
- Follow the assembly guide
Assemble the SO‑101 using the linked Assembly Guide.
- Optionally complete the assembly
If you bought a parts kit, you may optionally follow the Assembly Guide to assemble it.
- Assemble the SO-101 robot
Follow the assembly instructions on the SO-101 page under Robots in the sidebar.
- Align and Attach the Webcam
Push the webcam into the camera mount top while aligning the round joint holes.
- Secure the Webcam with an M2 Screw
Place the hex bolt in the hexagon-shaped hole, then screw an M2 screw into the opposite hole of the camera mount top’s round joint.
- Align the Camera Mount Halves
Push the camera mount top into the bottom, aligning their linear joint lines.
Software
- Install Miniforge
Download the Miniforge installer for your operating system and architecture, then run it.
- Create a Conda Environment
Create a conda environment named lerobot using Python 3.12.
- Create a UV Environment
As an alternative to conda, install Python 3.12 with uv and create a virtual environment using it.
- Activate the Conda Environment
Activate the lerobot conda environment every time you open a shell to use LeRobot.
- Install evdev in WSL Conda
When installing LeRobot in WSL with conda, also install evdev from the conda-forge channel.
- Activate the Environment on Linux
On Linux or macOS, activate the virtual environment before using LeRobot.
- Activate the Environment in PowerShell
In Windows PowerShell, activate the virtual environment before using LeRobot.
- Install evdev in WSL UV
When installing LeRobot in WSL with uv, install the evdev development library and the evdev Python package.
- Install FFmpeg on Ubuntu or Debian
Install the system-wide FFmpeg package with apt.
- Install FFmpeg on Apple Silicon macOS
Install FFmpeg with Homebrew on macOS running Apple Silicon.
- Install FFmpeg with Conda fallback
System-wide FFmpeg requires PyTorch >= 2.10 and TorchCodec >= 0.10. For older PyTorch versions, install FFmpeg from conda-forge with Conda.
- Clone the LeRobot repository
Clone the LeRobot repository and change into its directory.
- Install editable source extras with pip
Install the core scripts for recording, replaying, and calibration. The training extra installs policy-training dependencies, while all installs everything.
- Install editable source extras with uv
Use uv for the core scripts, training, and complete editable installations.
- Install the core LeRobot package
Install the base LeRobot package from PyPI with pip.
- Install core LeRobot with uv
Install the base LeRobot package from PyPI with uv.
- Add the required feature extras
Choose the feature-scoped extra matching the required workflow. When using uv, replace pip install with uv pip install.
- Install composite workflow extras
Install core_scripts for recording, replaying, and calibration. Combine core_scripts with training for recording and training, or install all extras.
- Install specific policy and hardware extras
Install extras for Pi0, SmolVLA, diffusion policies, simulation environments, or Feetech motors as needed.
- Install CUDA 12 PyTorch
Use the cu126 index when the driver supports only CUDA 12. This index installs torch 2.14; staying on CUDA 12 means using torch 2.14 or older.
- Install CPU TorchCodec
On Linux, install TorchCodec from the CPU index because PyPI provides a CUDA 13 build. For torch below 2.11, select the matching release from the TorchCodec compatibility table.
- Install LeRobot Package
Install LeRobot from the source tree with the all extra, or install the published package from PyPI.
- Install LeRobot With uv
Use uv pip with the cu126 backend to select the PyTorch index and install LeRobot in one step.
- Configure Locked CUDA Build
uv sync always installs the committed lock. To change its CUDA build, add the index under tool.uv.index and tool.uv.sources in pyproject.toml, then re-run uv lock.
- Install Build Dependencies
If build errors occur, install the listed Linux system dependencies with apt-get. These include cmake, build-essential, Python development headers, pkg-config, and FFmpeg libraries.
- Install the SO101 motor extra
Install the feetech extra for SO100, SO101, or Moss motors. Both dynamixel and feetech communicate with the motors through rustypot.
- Open the local Trackio dashboard
Trackio is included in the training extra. Local runs require no account or login, and their data is stored in a local SQLite database under ~/.cache/huggingface/trackio.
- Set the Space privacy option
New Spaces are public by default unless the organization’s default is private. Pass --trackio.private=true to create a private Space.
- Control console output capture
Trackio records the command that launched each run and its console output, including prints, logs, and tracebacks, on the run’s dashboard page. Pass --trackio.capture_console=false to keep console output out of the run.
- Generate a Hub write-access token
Generate a write-access token from the Hugging Face settings page.
- Authenticate the Hugging Face CLI
Run the CLI login command with the token in HUGGINGFACE_TOKEN and add it to the Git credential store.
- Store the Hugging Face username
Run the whoami command and store the displayed user name in HF_USER. The recording example uses this variable in the dataset repository ID.
- Configure the SO101 follower arm
Configure the SO101Follower with its serial port, follower ID, wrist camera, and top camera. Both OpenCV cameras use 640×480 resolution at 30 FPS.
- Configure the SO101 leader arm
Configure the SO101Leader with the specified serial port and leader-arm ID. The example uses /dev/tty.usbmodem5AB90689011 and my_leader_arm.
- Install the Feetech SDK
If LeRobot was installed from PyPI, run `pip install 'lerobot[feetech]'` instead. Otherwise, install the editable package with the Feetech extra using `pip install -e ".[feetech]"`.
- Visualize the SO100 URDF
Use rerun and the URDF visualizer plugin to visualize the SO100 URDF file. Run the command to open Simulation/SO100/so100.urdf.
- Visualize SO101 URDF Files
Use rerun and the URDF visualizer plugin to visualize the SO101 URDF files. Run the commands to open the new-calibration and old-calibration URDF files.
- Configure overhead camera software
Add the overhead camera in your software and set its resolution and FPS. Use 30 FPS, with a resolution of 640 x 480 or 1280 x 720 for a wider workspace. View the camera feed to confirm that the arm workspace is visible, and widen the screen resolution if needed to see more of the workspace; on Mac, QuickTime's New Movie Recording can also show the feed.
- Add the Overhead Camera in Software
Add the overhead camera in your software and set its resolution and FPS. The document recommends 640 × 480 resolution and 30 FPS even if the camera module supports higher or wider settings.
- Set Resolution and FPS
Set the resolution and FPS in your software. You will likely want 640 x 480 at 30 FPS because most models work with lower resolutions, while higher settings cause data bloat.
Calibrate and test
- Identify each arm’s USB port
Connect MotorBus to the computer with USB and power, then run `lerobot-find-port`. Disconnect the relevant leader or follower arm when prompted, record the corresponding port, and reconnect the USB cable.
- Grant Linux USB port access
On Linux, use the documented permission commands if access to the identified USB ports is required. Apply the commands to `/dev/ttyACM0` and `/dev/ttyACM1` as shown.
- Assign Unique IDs and Baudrates
Connect to each motor individually and assign it a unique ID. Configure the controller and all motors with the same baudrate; the settings are stored in EEPROM and only need to be configured once.
- Connect the Follower Controller
Connect the follower arm’s controller board to the computer with the USB cable and connect the power supply.
- Run the Motor Setup Command
Run the setup command for the so101_follower robot, replacing the shown port with the port found in the previous step.
- Set Up Motors Through the API
Create an SO101FollowerConfig with the follower port and an id that names the follower arm, then create the follower and call setup_motors().
- Connect Only the Gripper Motor
When instructed, connect the gripper motor alone to the controller board and press Enter. Do not daisy-chain it to another motor; the script then sets its ID and baudrate.
- Check Connections and Jumpers
If an error occurs, check the power supply, the USB cable between the computer and controller board, and the 3-pin cable from the controller board to the motor. On a Waveshare controller board, set both jumpers to the B channel (USB).
- Confirm the Gripper Motor ID
Check that the setup reports that the gripper motor ID has been set to 6 before continuing to the next motor.
- Connect the Wrist Roll Motor
Connect the controller board only to the wrist_roll motor and press Enter as instructed, so its ID and baudrate can be set.
- Set up each motor in sequence
Repeat the operation for each motor as instructed. Check the cabling at every step before pressing Enter, including the power supply cable.
- Complete the motor setup script
Run the setup process until the script finishes. The motors will then be ready to use.
- Repeat setup for the leader arm
Perform the same motor connection and setup steps for the leader arm.
- Configure the SO101 leader through the command
Run the motor setup command for the SO101 leader and paste the port found during the previous step.
- Configure the leader with Python
Create an SO101Leader with the specified port and id, then call setup_motors. The example uses /dev/tty.usbmodem585A0076841 and the id my_awesome_leader_arm.
- Position the Follower Joints
Move the follower robot so all joints are at the middle of their ranges.
- Run Follower Calibration
Run the follower calibration command with the follower port and a unique robot ID. The document also provides an API example using SO101FollowerConfig and SO101Follower.
- Move Every Follower Joint
After pressing Enter, move each joint through its full range of motion.
- Calibrate the Leader Arm
Repeat the follower calibration steps for the leader arm using the leader calibration command or API example.
- Calibrate the SO-101 leader arm
Connect the leader arm at /dev/tty.usbmodem58760431551 and identify it as my_awesome_leader_arm, then run calibration.
- Run leader calibration in Python
Create an SO101LeaderConfig with the specified port and id, connect the SO101Leader without automatic calibration, call calibrate(), and disconnect it.
- Validate using only camera images
Check that you can perform the task yourself using only the camera images.
- Review cadence before training
Read the per-episode and whole-session cadence reports before training on the dataset.
- Compare actual and target cadence
Compare achieved frequency with the target frequency. A session running at 25 Hz can still produce a dataset claiming 30 Hz, with motion that is simply faster than reality.
- Check pacing headroom bottlenecks
Check whether pacing headroom is near zero, which means the loop is saturated. In the example, observation used 74.4% of the work and had a worst time of 40.20 ms against a 33.3 ms budget.
- Override resumed job settings
Job settings come from the current command, so override --job.target, --job.timeout, and other settings as needed.
- Keep checkpoints on the Hub
Keep --save_checkpoint_to_hub=true so that the resumed run can itself be resumed later.
- Match the recorded robot setup
Use the same --robot.id as the recorded dataset so its calibration is found. Use the same --robot.cameras and task description.
- Check Platform Requirements
On Apple Silicon, use macOS 14 or newer for PyTorch and macOS 15 or newer for MPS acceleration; macOS 14 falls back to CPU. On Linux, the default PyTorch wheels require NVIDIA driver 580.65 or newer; older drivers require the PyTorch CUDA variant.
- Calibrate the SO-101 robot
Follow the calibration instructions on the SO-101 page under Robots in the sidebar.
- Calibrate Printer and Bed Level
Calibrate the printer and correctly set the print-bed level using the printer-specific instructions.
- Power On and View Video
Turn on the arm and view the feed from the newly installed camera. On Mac, you can also use QuickTime's New Movie Recording to see the feed.
- Adjust Camera Focus
The camera has manual focus and will initially look blurry. Twist the lens counterclockwise or clockwise until the video feed is clear.
- Check the Camera Video Feed
View the newly installed camera’s video feed; you should be able to see the workspace of the arms. On Mac, you can also use QuickTime’s New Movie Recording.
- Configure software and adjust focus
Data and training
- Record and upload five episodes
Adapt the example for the SO-101 follower and leader, their listed ports and IDs, and the front, wrist, and top OpenCV camera settings. Record 5 episodes for the task "Grab the black cube," display the data, and upload the dataset to ${HF_USER}/record-test with streaming encoding and 2 encoder threads.
- Create the LeRobot recording dataset
Convert the robot's action and observation features into dataset features, then create a LeRobotDataset with the specified repository ID and 30 FPS. The example enables videos, sets the robot type to the follower's name, and uses four image writer threads.
- Connect the robot and teleoperator
Initialize the keyboard listener and rerun visualization, then connect the follower robot and SO101 leader teleoperator. Create the default teleop-action, robot-action, and robot-observation processors.
- Record five teleoperated episodes
Run the recording loop until five episodes are recorded or the stop-recording event is set. Each episode uses 30 FPS, a 60-second control time, the task description "My task description", and enabled data display.
- Reset the Environment Between Episodes
When recording has not stopped and another episode is needed, reset the environment with record_loop using the configured FPS, processors, teleoperation, control time, and task description.
- Re-record the Current Episode
When re-recording is requested, clear the episode buffer, reset the re-record and early-exit events, and continue to the next recording iteration.
- Save Each Recorded Episode
Save the completed episode to the dataset and increment the episode index.
- Finalize the Recorded Dataset
Finalize the dataset after recording all episodes.
- Disconnect Devices and Push Dataset
Stop recording, disconnect the robot and teleoperation devices, and push the dataset to the Hub.
- Keep Camera Names Consistent
The wrist and top camera names become observation.images.wrist and observation.images.top. A deployed policy expects the same camera names and resolution, so reuse the same --robot.cameras value throughout the tutorial.
- Use the Timestamped Dataset Name
lerobot-record appends the session date and time to --dataset.repo_id, producing a unique name such as record-test_20260929_143012.
- Find the Full Dataset Name
Find the complete timestamped dataset name on the Hugging Face profile or by listing the local LeRobot cache directory.
- Disable Automatic Dataset Timestamping
Add --dataset.no_stamp=true to keep the exact dataset repository ID passed to lerobot-record.
- Locate the Local Dataset
The local dataset is stored under ~/.cache/huggingface/lerobot/{repo-id}.
- Upload the Dataset to the Hub
At the end of data recording, the dataset is uploaded to the user's Hugging Face page.
- Print the Uploaded Dataset URL
Print the Hugging Face dataset URL using the timestamped dataset name.
- Tag the Dataset for Discovery
The dataset is automatically tagged with LeRobot, and custom tags such as tutorial can be added with --dataset.tags.
- Push the Local Dataset Manually
The local dataset can also be pushed to the Hugging Face Hub manually.
- Upload the recorded dataset
Upload the local recording to a Hugging Face dataset repository.
- Store recordings in LeRobotDataset format
Record robot-operation data with lerobot-record; the data is stored on disk in the LeRobotDataset format during recording.
- Disable automatic dataset uploading
Start recording with --dataset.push_to_hub=False when the dataset should not be uploaded.
- Allow recording checkpoints
Begin recording and let lerobot-record create checkpoints automatically during the recording session.
- Resume an existing recording
Re-run the same command with --resume=true, pass the full timestamped dataset name with --dataset.repo_id, set --dataset.num_episodes to the number of additional episodes, and set --dataset.root to the required local path.
- Start a new timestamped dataset
Re-run the command without --resume=true to create a new dataset with a new timestamp. If using --dataset.no_stamp=true, delete the existing dataset directory first.
- Configure recording timing and episodes
Set --dataset.episode_time_s=60 for each episode, --dataset.reset_time_s=60 for environment resets, and --dataset.num_episodes=50 for the total number of episodes.
- Control recording with keyboard shortcuts
Press Right Arrow or n to stop the current episode or reset time and move to the next; press Left Arrow or r to cancel and re-record the current episode; press Escape or q to stop the session immediately, encode videos, and upload the dataset.
- Use terminal recording controls
On Wayland, headless systems, or macOS without the required permission, launch lerobot-record from an interactive terminal and keep the terminal focused to use n, r, and q.
- Choose a grasping and bin-placement task
Use grasping an object at different locations and placing it in a bin as the starting task.
- Record at least fifty episodes
Record at least 50 episodes and allocate 10 episodes to each object location.
- Keep recordings visually consistent
Keep the cameras fixed and use consistent grasping behavior throughout the recordings.
- Keep the object visible
Make sure the manipulated object remains visible on the cameras.
- Proceed to neural network training
After collecting the dataset, proceed to the neural network training sections.
- Increase variation gradually
Do not add too much variation too quickly because it may hinder the results.
- Open the first episode from its printed link
Enter the repo id on the visualizer's home page, or use the printed link to open episode_0 directly.
- Sign in to view a private dataset
Sign in to the visualizer when the dataset is private.
- Replay an episode with the command
Run the replay command with the so101_follower robot type, its port, a robot ID, and the dataset repository ID and episode number.
- Replay an episode with the API
Set episode_idx to 0, configure and connect SO101Follower, load episode 0 from the dataset, select its action column, send each action to the robot, sleep for the dataset frame timing, then disconnect.
- Start ACT policy training
Run lerobot-train with the recorded SO-101 dataset, ACT policy, output directory, job name, CUDA device, Trackio, and policy repository.
- Select the recorded dataset
Set --dataset.repo_id to your dataset’s full name, including its timestamp.
- Configure the ACT policy
Set policy.type=act. ACT loads configuration_act.py and automatically adapts to the dataset’s number of motor states, motor actions, and cameras, such as wrist and top cameras.
Use it
- Try the Colab notebooks
Try the LeRobot Colab notebooks.
- Keep the LeRobot cheat sheet
Keep the cheat sheet at hand; it lists every LeRobot command in one place.
- Restore recording controls in an interactive terminal
On Linux, run lerobot-record in a focused interactive TTY. Use the arrow keys or Escape, or the n / r / q equivalents; no $DISPLAY setup is needed.
Common problems
- Key components and regional replacement links are unavailable or uncertain #68 #16 #76 #6 #161
- Servo and clamp clearances are too tight in printed base parts #63 #12
- Leader arm walls fall below the print service minimum #175
- Mount helper orientation is poorly suited to standard printing #119
- Loose base and shoulder joints cause persistent vibration #24
- Power-supply voltage compatibility with the servo controller is unclear #142
- Calibration fails without correct servo positions or physical stops #60 #170
- URDF joint limits, axes, and dynamics are incorrectly defined #54
Datasets
Parts
- 12 STS3215 7.4V servos for a pair (6 per follower; leader mixes 1/345, 1/191 and 1/147 gear ratios)
- 2 motor control boards, 2 USB‑C cables, 2 power supplies
- 4 table clamps
- 3D-printed structure (STL published; PLA+ at 15% infill recommended)
Tools
- 3D printer or a printing service
- Phillips screwdrivers #0 and #1
- Putty knife for removing prints
Skills
- Basic Python and command line
- LeRobot setup and calibration
Software
- LeRobot (Hugging Face)