so-arm100

Isaac Lab external project for SO-ARM100/101 arm robot.

Open-source 3D-printed parallel gripper for SO-ARM100, SO-100, SO-ARM101, SO-101, SO-ARM-102 robot arms. Feetech STS3215, 84 mm stroke, camera mounts, STL files, BOM and assembly guide.

An open-source SO-ARM102 robot arm, successor to SO-ARM100 and SO-ARM101. LeRobot-compatible leader-follower kit for teleoperation and imitation learning, with improved rigidity, parallel gripper, STEP/STL files and BOM.

Custom firmware for Feetech STS32XX servos (GD32F130) in SO-ARM100, SO-ARM101 and SO-ARM102 robot arms, with smooth motion, tunable PID and gravity compensation.

HALO is a live multimodal robot agent for real-time manipulation. It combines natural voice interaction, scene understanding, narrated execution, and autonomous skill orchestration in a robotics runtime designed for continuous control.

ROS 2–based SO-ARM100 Reinforcement Learning System

ROS 2 package for SO-100 and SO-101 arms

Real-time, interactive robot arm simulation you can command over the network. Python physics simulator, Go gRPC relay, and a Bubble Tea TUI, orchestrated with Docker Compose.

SO-101 (SO-ARM100) web control panel: slider control + trajectory record & playback, powered by LeRobot. Apache 2.0.

WebXR teleoperation for LeKiwi (SO-101 arm + omni base) and other SO-101 robots from a Meta Quest headset — closed-form kinematics, no APK, no Unity

From-scratch RGB-D perception-to-grasp pipeline in MuJoCo: segmentation -> point cloud -> 6-DoF pose (yaw-sweep + trimmed ICP, 0.58mm) -> grasp -> closed-loop pick (20/20)

Inverse kinematics and a safe driver for the SO-101 arm (LeRobot) in C: RobotKinematics 3.6-41x faster, Python binding, ESP32-ready

Sim-to-real for the LeRobot SO-101 arm: MuJoCo scene, scripted expert, ACT policy trained on RunPod from simulated demos only, rolled out on the real robot. Task: pick a stress ball and place it on a toilet paper roll, wrist camera only. Work in progress.

Lightweight per-camera scalar gating for multi-view robot imitation learning — plug-in module for LeRobot ACT, no stage annotations required.

SO-101 robot arm in MuJoCo: keyboard teleop -> recording -> LeRobot dataset -> ACT training

하드웨어 없이 배우는 피지컬AI — SO-ARM100 로봇팔을 MuJoCo와 브라우저 3D에서 움직이는 실습실 (Python API · Claude Code MCP)

MuJoCo pick-and-place for the SO-101 arm with an evaluation harness tested against its own failure modes: a contact gate that rejects faked grasps, paired ablations with pre-registered predictions, and an operating envelope that shows competence is an annulus, not a success rate.

SmolVLA (450M VLA) fine-tuned for language-conditioned pick-and-place on free-tier compute: 0% -> 55% -> 90%, incl. a cross-architecture data study vs ACT

Fine-tune NVIDIA Isaac GR00T N1.7 (3B VLA) on a single 24 GB RTX 4090 without root: patches, install script, MuJoCo evaluation, open datasets, 1800 evaluated attempts as JSON

End-to-end imitation learning on a MuJoCo SO-ARM100 work-cell: scripted expert -> LeRobotDataset -> ACT -> domain-randomization robustness study (90% nominal, 87-100% under DR)