🦾 SO-101 arm and LeRobot

O LeRobot is an open library from Hugging Face, built with Python and PyTorch. It brings together everything you need to teach a robot by demonstration: control motors, read cameras, record episodes, train policies, and run them on the real arm. The SO-101 is the project’s reference arm: six servos on each arm, parts you can print or buy ready-made, and hobby-level cost. With the leader/follower pair, you demonstrate the task by moving the leader, and the follower repeats it.

How to read this map: The circles are the joints (five plus the gripper at the end). The cyan arm is the leader: you hold it, and it only measures positions. The teal arm, shown glowing, is the follower: it has motors and copies the pose. The computer in between reads one and sends commands to the other, dozens of times per second.

Why learn this

The SO-101 and LeRobot combination is currently the shortest and cheapest path from “I have an arm” to “the arm has learned a task.” Everything you do here (ports, IDs, calibration) is the foundation for modules 5.4, 5.5, and 5.6.

Key concepts

Every STS3215 comes from the factory with the same ID. Since the six servos on an arm share the same wires, each one needs a different ID before final assembly: from 1 at the base to 6 at the gripper. Once assembled, calibration teaches the software where the midpoint and limits of each joint are, so the leader and follower speak the same position “language.”

How to read this map: The teal line is the bus: one data wire, plus power, that runs through all the servos in a chain. The number in each box is the ID; below it is the corresponding joint. If two servos have the same ID, both respond together and the arm goes haywire. That’s why you configure them one at a time.

Step by step

Goal: the commands for steps 1, 2, and 4, using the syntax that worked in the INEMA robotica-real2sim guide with LeRobot 0.6.1 (Oct 2026). Flags vary between versions: check --help and the current documentation.

source ~/lerobot-env/bin/activate        # the environment from topic 3
lerobot-find-port                       # disconnect and reconnect when prompted: this shows the port

# servo IDs (one servo at a time on the cable; follow the on-screen instructions)
lerobot-setup-motors --robot.type=so101_follower --robot.port=/dev/ttyACM0
lerobot-setup-motors --teleop.type=so101_leader  --teleop.port=/dev/ttyACM1

# calibration (arm assembled; id names the calibration file)
lerobot-calibrate --robot.type=so101_follower --robot.port=/dev/ttyACM0 --robot.id=seguidor
lerobot-calibrate --teleop.type=so101_leader  --teleop.port=/dev/ttyACM1 --teleop.id=lider

How to read this map: The gray arc shows how far the servo can turn; teal shows the range allowed by the assembly before one part hits another. Calibration measures the teal range and the midpoint (cyan line). That way, “50% of the elbow” means the same pose on the leader and follower, even if their servos were assembled at different angles.

Why learn this

Almost every arm problem where it “shakes,” “moves the wrong way,” or “doesn’t copy properly” comes from a swapped ID or poor calibration. Taking your time with this step saves hours of troubleshooting during the training in module 5.4.

Key concepts

LeRobot is installed with pip, inside its own virtual environment (it includes PyTorch, which is large and has its own version requirements). The feetech extra installs support for the SO-101 servos; you’ll need additional extras for training in 5.4. The official documentation also covers installation with conda and ffmpeg (used to record video of the episodes). The names of extras and the required Python version can change: check the installation page before you begin.

Goal: create the environment, install LeRobot with Feetech servo support, and grant access to the USB port.

python3 -m venv ~/lerobot-env
source ~/lerobot-env/bin/activate
pip install --upgrade pip
pip install "lerobot[feetech]"           # check the current documentation for the version and extras
pip show lerobot                        # shows the installed version
lerobot-find-port --help                # the lerobot-* commands should be available

# Linux: permission to use USB ports without sudo (log out and back in afterward)
sudo usermod -aG dialout $USER

Why learn this

AI libraries for robots change quickly. Knowing how to isolate the installation, check the version, and read the --help is what makes the difference between a project that runs again six months from now and one that breaks with the first update.

Key concepts

During teleoperation, the computer repeats a short loop: it reads the six joint positions of the leader, converts them using the calibration, and sends those same positions as targets to the follower. This is what will happen when recording demonstrations in module 5.4, except nothing is saved. So before recording, practice your movements here: smooth, without jerks, and always within the clear area.

How to read this map: follow the teal arrows from left to right and return along the dashed cyan arrow: it's a continuous loop while the command is running. If the follower seems "delayed," the loop is slow (cable, busy computer, or demanding camera); if it moves in the wrong direction, the problem is in the middle box: the calibration.

Goal: your first teleoperation, without cameras. Use the ports and the id from topic 2. Before pressing Enter, put the leader in a pose similar to the follower's: at startup, the follower moves to the leader's pose.

lerobot-teleoperate \
  --robot.type=so101_follower --robot.port=/dev/ttyACM0 --robot.id=seguidor \
  --teleop.type=so101_leader  --teleop.port=/dev/ttyACM1 --teleop.id=lider
# Ctrl+C ends the program (it doesn't replace the hardware stop switch: see topic 6)

Why learn this

The policy in module 5.4 learns to imitate exactly what you do during teleoperation, including jerks and hesitations. A calm, consistent hand here is worth more than a larger model later.

Key concepts

The policies in module 5.4 see through the cameras, so the setup needs to stay stable: fixed cameras, consistent lighting e simple background. A common setup uses two USB cameras: one in front (or overhead), which sees the whole table, and another on the wrist or nearby, which gets a close-up of the gripper. Anything that changes between recording and use (the lamp, the tablecloth, the camera angle) can confuse the policy.

How to read this map: Seen from above, the arm (highlighted) sits at the edge of the table; the semicircle shows how far it can reach, and the task objects sit within it. The two cyan cameras face this area. The dashed red outline marks the clear area, and the hardware stop switch sits outside it, near you.

Goal: find the index of each camera and teleoperate while viewing its image alongside the joint positions.

lerobot-find-cameras opencv           # lists the cameras and saves a photo from each (check the index)

lerobot-teleoperate \
  --robot.type=so101_follower --robot.port=/dev/ttyACM0 --robot.id=seguidor \
  --robot.cameras="{ frente: {type: opencv, index_or_path: 0, width: 640, height: 480, fps: 30}}" \
  --teleop.type=so101_leader --teleop.port=/dev/ttyACM1 --teleop.id=lider \
  --display_data=true

How to check: opens a window showing the "front" camera image and the joint positions in real time while you teleoperate. For the second camera, add another entry to the dictionary (for example lateral: {type: opencv, index_or_path: 2, ...}) with the index that find-cameras showed. Syntax checked against LeRobot 0.6.1; check your version.

Why learn this

In learning by demonstration, the camera is the main sensor. A repeatable workspace makes the difference between a policy that gets it right almost every time and one that only works on the afternoon it was recorded.

Key concepts

The SO-101 is small, but strong enough to pinch a finger in its gripper, knock a glass of water onto a laptop, or hit someone in the face if they lean over the table. The risks increase when a trained policy, rather than your hand, is in control (module 5.4): it may make a move you did not expect. The rule is the one from module 5.1: the stop does not depend on software. A switch in series with the servos’ power supply, within reach, cuts power immediately. When power is cut, the arm loses its support and falls onto the table, so keep anything fragile out from underneath it.

Before each session

Why learn this

The safety habits you build with a small arm are the same ones used in industry with large arms: marked areas, a physical stop, and a test before operation. Anyone taking the arm to a school or fair needs these habits from day one.

Key concepts