LeRobot SO-100M: The Ultimate Guide

LeRobot SO-100M Hero Image

The SO-10xARM is a cutting-edge, fully open-source robotic arm project launched by TheRobotStudio. Integrated with Hugging Face's LeRobot, it aims to lower the barrier for real-world robotics using PyTorch and imitation learning.

What's New in SO-ARM101

Wiring Optimization

Improved design prevents disconnection at joint 3 and removes motion limits.

Optimized Ratios

New motor gear ratios eliminate the need for external gearboxes.

Real-time Follow

Leader arm follows follower in real-time, crucial for human intervention training.

Hardware & Specification

The system relies on Feetech ST-3215 series bus servos. The "Arm Kit Pro" version supports 12V operation for higher torque on the follower arm.

Component Follower (SO101) Leader (SO101)
Joint 2 ST-3215-C001 (1:345) ST-3215-C001 (1:345)
Joints 1 & 3 ST-3215-C044 (1:191) ST-3215-C044 (1:191)
Joints 4, 5, 6 ST-3215-C046 (1:147) ST-3215-C046 (1:147)
Operating Voltage 12V (Pro) / 5V (Std) 7.4V / 5V
Important: The leader arm always uses 7.4V motors. Plugging a 12V supply into the leader arm bus may burn your motors. Always verify your IDs (F1-F6 and L1-L6) before power-up.

Initial System Environment

Ubuntu x86

  • Ubuntu 22.04 LTS
  • CUDA 12+
  • Python 3.10
  • Torch 2.6+

Jetson Orin

  • JetPack 6.0 / 6.1
  • Python 3.10
  • Torch 2.3+

3D Printing Guide

Step 1

Settings: PLA+, 0.4mm nozzle, 0.2mm layer, 15% infill.

Step 2

Preparation: Level bed, clean with IPA. Supports everywhere > 45°.

Step 3

Print: Source files via Makerworld.

Install LeRobot

It's recommended to use a verified stable repository for consistent performance.

Bash
# Clone Seeed stable version
git clone https://github.com/Seeed-Projects/lerobot.git ~/lerobot
cd ~/lerobot

# Install dependencies (feetech support)
pip install -e ".[feetech]"
conda install ffmpeg=7.1.1 -c conda-forge

Motor Configuration

Connect only one motor at a time for ID assignment. Use 5V for leader motor calibration and 12V for follower (if Pro version).

Bash
# Find arm ports
lerobot-find-port

# Setup Follower (e.g., port /dev/ttyACM0)
lerobot-setup-motors --robot.type=so101_follower --robot.port=/dev/ttyACM0

# Setup Leader (e.g., port /dev/ttyACM1)
lerobot-setup-motors --teleop.type=so101_leader --teleop.port=/dev/ttyACM1

Calibration

Manual calibration ensures both arms share the same physical-to-digital mapping.

Bash
# Set permissions
sudo chmod 666 /dev/ttyACM*

# Calibrate Follower
lerobot-calibrate --robot.type=so101_follower --robot.port=/dev/ttyACM0 --robot.id=follower_arm

Data Collection

Record 50+ episodes for a stable ACT policy. Aim for 10 episodes per location variation.

Bash
# Record locally
lerobot-record \
  --robot.type=so101_follower --robot.port=/dev/ttyACM0 \
  --teleop.type=so101_leader --teleop.port=/dev/ttyACM1 \
  --robot.cameras="{ front: {type: opencv, index_or_path: 0, width: 640, height: 480}}" \
  --dataset.repo_id=my_username/test_task \
  --dataset.num_episodes=50 \
  --dataset.push_to_hub=false

Training & AI Policy

Train an ACT (Action Chunking with Transformers) policy. On an RTX 3060, 50 datasets take ~6 hours to reach optimal loss.

Training
lerobot-train \
  --dataset.repo_id=my_username/test_task \
  --policy.type=act \
  --output_dir=outputs/train/act_test \
  --steps=300000 \
  --policy.device=cuda

Troubleshooting (FAQ)

Check cable connections and ensure correct voltage (5V/12V). Verify the bus is powered.

Servo zero-offset is abnormal. Power cycle the robot. If persists, use the RoboController tool for middle-position calibration (set to 2048).

Grant serial permissions using: sudo chmod 666 /dev/ttyACM*

Citations & Resources

This guide is based on the official Seeed Studio and Hugging Face documentation. For latest updates, visit:

Back to Projects