The SO-101 (SO-ARM101) specs

BOM by category · USD from priced lines

Offers in CN · EU · GLOBAL · US from 8 vendors

Subsystems · assembly groups

Fabrication: FDM print ×17

Firmware

Targets

Feetech STS3215 servo (EEPROM: ID 1-6 and bus baudrate, written by lerobot-setup-motors), Waveshare / Seeed serial bus servo driver board (USB serial bridge, no user firmware)

Repo

github.com/huggingface/lerobot

Calibrate

lerobot-calibrate --robot.type=so101_follower --robot.port=<port> --robot.id=<id>, then lerobot-calibrate --teleop.type=so101_leader --teleop.port=<port> --teleop.id=<id>. Middle-of-range pose, Enter, sweep all joints except wrist_roll, Enter.

Notes

No custom firmware: servos keep Feetech firmware and LeRobot (lerobot.robots.so_follower, lerobot.teleoperators.so_leader) talks to them over the serial bus. so100_follower and so101_follower are the same config class in LeRobot 0.6.x. Optional debug tools: Feetech FD software (Windows, https://www.feetechrc.com/software.html) or FT_SCServo_Debug_Qt (Ubuntu). The repo also has a WebUI 3-point calibration guide (Software/WEBUI_CALIBRATION.md) that references a pi_servo_studio.py script and a --webui flag not found in the repo or in lerobot 0.6.2.

Policies

• Generalist SO-101 pick and place (world action model, fine-tune with the included LoRA recipe) · LeRobot (Flux3Policy) · huggingface.coBlack Forest Labs FLUX 3 Action, 7B. Trained on the SO-101 episodes of lerobot/community_dataset_v3. Two cameras (scene + wrist), 6-D joint-delta actions with absolute gripper, 30 Hz. FLUX Kommunity License: non-commercial use unless you are a qualifying user; needs a large GPU.
• SO-100/101 language-conditioned manipulation (VLA, zero-shot or fine-tune) · LeRobot (MolmoAct2) · huggingface.coLeRobot port of allenai/MolmoAct2-SO100_101, fine-tuned by Ai2 on an SO-100/101 mixture with absolute joint-pose control. Cameras map to cam0/cam1 via --rename_map. The processor includes the joint sign/offset fix for the pre-0.5 calibration convention. No license field on either model card.
• SmolVLA base (pretrained VLA to fine-tune on your own SO-101 dataset) · LeRobot (SmolVLA) · huggingface.co450M. Per the SmolVLA paper (arXiv 2506.01844), pretrained on community LeRobot datasets that are mostly SO-100 recordings; not meant to run zero-shot. Fine-tune with `lerobot-train --policy.path=lerobot/smolvla_base`. Apache-2.0.