Awesome SO-ARM

A curated list of software, hardware, simulators, policies, and practice guides for the SO-ARM family (SO-ARM100 / 101 / 102) — the open-source arm of the LeRobot ecosystem.

English · 简体中文

---

The SO-ARM design by TheRobotStudio has become the de-facto open hardware platform for imitation-learning research: low-cost bus servos, 3D-printable structure, first-class support in LeRobot, and a rapidly growing variant ecosystem. This list tracks what exists, what works, and what to watch out for.

Contributions welcome — see CONTRIBUTING.md. Star counts are snapshots (not live).

Official & Core

• TheRobotStudio/SO-ARM100 — the original open arm: BOM, 3D files, assembly guides. Everything else descends from here.
• huggingface/lerobot — reference framework: calibrate / teleoperate / record / train toolchain with SO-ARM as a first-class citizen.
• roboninecom/SO-ARM-102 — next-gen kit: 5-DOF + gripper, PET-CF frame, parallel gripper.

Hardware & Upgrades

• roboninecom/SO-ARM100-101-Parallel-Gripper — 3D-printed parallel gripper for STS3215-class servos; a step change in grasp capability over the stock single-servo finger.
• brukg/SO-100-arm — v1.3 C++ firmware line.
• Servo batch notes: compatible clones (e.g. HX-series) speak the Feetech protocol but may ship at non-default baudrates — scan before you assume 115200.

Simulation & ROS

• ros-physical-ai/ros2_so_arm — ROS 2 description + drivers + MoveIt config for SO-ARM100/101.
• MuammerBay/isaac_so_arm101 — Isaac Lab RL tasks and tutorials for SO-ARM (URDF → OpenUSD pipeline).
• mujocolab/mjlab — Isaac Lab API on MuJoCo Warp; increasingly the default free/GPU-accelerated choice for arm+locomotion RL (no NVIDIA stack lock-in).

Policies, Training & Data

• lerobot/smolvla_base — 450M VLA pretrained incl. SO-100 data; benchmark-finetuned siblings (smolvla_libero, _metaworld, _robotwin) on the same org.
• aalmuzairee/squint — visual SAC with sim-to-real on a real SO-101: reach/lift/place/stack solved in minutes of training.
• MuammerBay/SO-ARM101_MoveIt_IsaacSim — earlier MoveIt integration (superseded by the Isaac Lab repo above).
• Datasets: search so100/so101 on the lerobot org for community teleop datasets usable for pretraining.

Agents & Harnesses (emerging)

The VLM-agent / harness niche for SO-ARM is still sparse — most robot-use agent work targets other platforms. Related efforts:

• jixinyan/Embodied-DeepSeek-Harness — composable agent teams with pluggable robot interfaces (TypeScript, plugin architecture worth studying).
• vicfei/so_arm_agent — planner / primitives / executor / verifier harness for SO-ARM101: LLM as low-frequency consultant, deterministic primitives with a code-veto supervisor, mock-first development.
• If you are building an LLM/VLM harness that drives an SO-ARM, please open a PR.

Practice Guides (community)

• SO-ARM101: from unboxing to first dataset on a WSL2 laptop — no-sudo WSL2, USB-serial via TCP bridge, Feetech-SDK patching, calibration, scripted-expert data collection, MJPEG camera streaming; includes a full troubleshooting log (raw-bytes-first rule, COM handle hygiene, incremental saves).
• Robot Use Agent roadmap for SO-ARM — planner/primitives/executor/verifier harness design with code-veto safety layering.

Ecosystem Landscape

• Robot harness ecosystem snapshot (2026-10) — five-repo survey situating SO-ARM agent tooling in the broader harness wave.
• Locomotion & world-model snapshot (2026-10) — microduck stack + SuperMimic; the low-compute training routes relevant to arm RL too.

License

CC0-1.0