Hardware specification
Low-cost. Precise enough
for real policy work.
Architecture
Leader + follower pair
DoF
6 per arm
Follower motors
6× STS3215 (1/345 gearing)
Servo torque
30 kg·cm at 12V
Feedback
360° magnetic encoder
Vision
Wrist + external camera
Compute
NVIDIA Jetson Orin compatible
Build cost
From ~$130 (DIY 3D print)
License
Apache 2.0 (Hugging Face)
Software stack
The full Hugging Face
embodied AI toolkit.
LeRobot (Hugging Face)Core framework
ACT policyAction chunking
Diffusion PolicyBehaviour cloning
PyTorchTraining backend
Hugging Face HubDataset + model sharing
Feetech SDKMotor control
How it learns
Demonstration to
autonomous policy.
A human operates the leader arm. The follower mirrors every movement. Cameras record the scene. The full episode — joint positions, gripper state, camera frames — is saved as a dataset episode.
Repeat for 50–200 episodes. Train an ACT or diffusion policy on those demonstrations. Deploy the policy to the follower arm. It runs autonomously.
Typical task: master in minutes, train in 4 hours.
Arm variants
SO-101 versus the
B601 — when to use which.
SO-ARM101
PoC & research
Low cost. Fast setup. Native LeRobot support. Ideal for demonstrating feasibility on a defined task before committing to industrial hardware.
reBot B601
Pre-production cell
Industrial actuators. ±0.2mm repeatability. Larger reach and payload. For policies moving to a real production environment.