사용 방법
- 설치 명령을 복사하고 자신의 에이전트 환경에 설치하기 전에 소스 패키지를 확인하세요.
- 필요한 도구, 인증 정보, 지원 환경은 원본 문서에서 확인하세요.
- 입력과 원하는 산출물을 설명하고 출처와 결과를 대조하세요.
npx skills add https://github.com/google-deepmind/robotics-antigravity-plugin --skill so101-bringup시작 전 확인
- 설치에는 모델 구독, API 한도, 파일 접근 권한이 포함되지 않습니다.
- 별점 수와 설치 수는 활동 지표이며 독립적인 품질·보안 검증이 아닙니다.
원본 저장소의 제목, 예시, 코드, 표, 링크, 이미지를 유지해 표시합니다.
SO-101 Hardware Bring-Up, Camera Verification & Stage 3 "Hello World"
Follow this staged workflow under R1–R16 (auto-advancing across software, --dry-run, and read-only steps without extra confirmation prompts) to bring up an SO-101 follower arm (so101_follower, 6× Feetech STS3215 servos) and 1 or more cameras (minimal setup: arm + 1 camera; recommended setup: fixed/overhead + wrist camera) using safety_rules.md, hardware_limits.md, and tabletop_pick_place.json. The fast path to Hello World is Stage 0 → 1 → 2 → 3; optional Stage 4 runs afterwards (required before Stage 5):
- Stage 0 — Environment, `NOTES.md` & Dedicated `tmux` Session `robot` (`R11–R16`, auto-advances to Stage 1):
- Summarize safety and working rules
R1–R16from safety_rules.md before touching hardware. - Install with
uvand shallow clones. LeRobot provideslerobot-find-port,lerobot-calibrate, motor drivers, and forward/inverse kinematics (RobotKinematics,kinematicsextra); SO-ARM100 is only needed for the URDF and its meshes, so sparse-checkoutSimulation/SO101. Checkffmpeg, and initializeNOTES.mdand git (R12,R13):
uv venv .venv && source .venv/bin/activate
git clone --depth 1 https://github.com/huggingface/lerobot repos/lerobot
uv pip install -e "repos/lerobot[feetech,kinematics]" opencv-python numpy
git clone --depth 1 --filter=blob:none --sparse https://github.com/TheRobotStudio/SO-ARM100 repos/SO-ARM100
git -C repos/SO-ARM100 sparse-checkout set Simulation/SO101- Create a persistent
tmuxsession namedrobot(tmux new-session -d -s robot), verify two-way command/capture (tmux send-keys+tmux capture-pane), tell the human how to watch live (tmux attach -t robot), and proceed directly to Stage 1 without waiting for sign-off.
- Stage 1 — Read-Only Serial Bus & Camera Discovery (Torque OFF, `R4`, `R16`):
- Auto-detect the serial port with read_motors.py
--port auto (uses the single /dev/tty.usbmodem* or /dev/ttyACM* device) to verify all 6 STS3215 servos (IDs 1..6 at 1,000,000 baud) respond with Torque_Enable = 0 and read their EEPROM calibration into LeRobot's default calibration cache when valid (calibration_valid: true and calibration_file in motors_readonly.json). Only if zero or several candidates are found, fall back to lerobot-find-port inside tmux session robot (human unplugs/re-plugs USB):
python3 scripts/read_motors.py --port auto --output data/stage1/motors_readonly.json- Automatically enumerate OS cameras, capture stills (
camera_0.png,camera_1.png, ...), and check lighting brightness (0–255, flaggingTOO_DARKorTOO_BRIGHT_OR_GLARE) using discover_cameras.py. If exactly 1 camera is found, use it without asking (rolewristif the jaw tips are visible at the bottom of the still, otherwisefixed). Ask the human about roles only when 2 or more cameras are found or a laptop built-in camera is detected (not recommended; it must stay completely still facing the robot). - Save the port,
calibration_file(andrest_posewhen computed from
EEPROM), camera roles/indices, and stills in config/robot_config.json and NOTES.md.
- Stage 2 — Joint Calibration, `rest_pose` & Abort Key (`R1–R5`, single `"go"` for bring-up motion):
- Reuse motor EEPROM calibration by default without asking the human to
recalibrate: if data/stage1/motors_readonly.json reports calibration_valid: true (or a valid calibration file already exists in LeRobot's cache and matches EEPROM via run_preflight()), reuse calibration_file and skip lerobot-calibrate. Otherwise (uncalibrated EEPROM or explicit recalibration request), with torque OFF (R16), run lerobot-calibrate inside tmux session robot (human moves arm to middle of joint ranges + Enter, then sweeps full min-to-max ranges + Enter) and flag any body joint with span < 45° (or gripper < 20°):
lerobot-calibrate --robot.type=so101_follower \
--robot.port=/dev/tty.usbmodem... --robot.id=my_follower- Automatically run
--mode abort-test --dry-run(R2, torque OFF, planned from the Stage 1 read-only snapshot) without asking for confirmation, then ask the human once for"go"(R1, covering--mode abort-testand Stage 3hello_world.pywhile the workspace remains clear; anyconnect()turns torque ON, so nothing connects before the"go"). Run a single slow5°base nudge with check_robot.py during which the human typesq+Enter; the script recordsrest_posebefore moving:
python3 scripts/check_robot.py --mode abort-test --dry-run
python3 scripts/check_robot.py --mode abort-test- Confirm that
q+Entertriggers an immediate smooth unwind alongtravelled_waypointsback torest_pose(R3) and setsTorque_Enable = 0(R4), then proceed directly to Stage 3.
- Stage 3 — "Hello World" Motion & Video (Minimal-Setup Milestone, `R1–R8`, auto `--dry-run` + covered by Stage 2 `"go"`):
- Works with the minimal setup (arm + 1 camera, e.g., wrist camera) or the 2-camera setup using hello_world.py (or
check_robot.py --mode hello-world). Its built-in pre-flight (R6) verifies all 6 servos, the calibration, and<= 0.30 sframe freshness on every configured camera before torque ON. - Run
--dry-runautomatically without asking (R2), then execute livehello_world.pyunder Stage 2's powered-motion"go"(only re-asking underR1if pre-flight failed or the human entered the workspace): start timestamped MP4 video recording (logs/stage3_hello_world_wrist.mp4,R7) before leavingrest_pose, move smoothly fromrest_poseto the calibration midpoint (`0.0°` on all 5 body joints) withinR5limits (<= 45°/step,<= 90°/s,5°margin, FK height floor), open and close the gripper jaws once, return smoothly torest_pose, and disable motor torque (R4):
python3 scripts/hello_world.py --config config/robot_config.json --dry-run
python3 scripts/hello_world.py --config config/robot_config.json --output-mp4 logs/stage3_hello_world_wrist.mp4- Inspect start, midpoint, and end video frames before reporting Stage 3 with the 6-field
R14report.
- Stage 4 (Optional: After Hello World, Required Before Stage 5) — USB Topology/Bandwidth, Reference Frames & 10-Second Video (`R6`, `R7`, `R10`):
- Automatically inspect the OS USB tree (
system_profiler SPUSBDataType/lsusb -t), compareYUYVvsMJPGbandwidth against USB 2.0 (480 Mb/stheoretical /~320 Mb/susable), and benchmark live streaming FPS and frame gaps across all configured cameras using benchmark_cameras.py:
python3 scripts/benchmark_cameras.py --config references/tasks/tabletop_pick_place.json- Automatically check lighting/glare diagnostics and save baseline reference frames for drift detection (
R10) with check_cameras.py, then record a 10-second timestamped MP4 + JSONL stream with record_overhead.py. If a fixed/overhead camera is configured, proceed to optional Stage 5 in ../so101-calibration/SKILL.md:
python3 scripts/check_cameras.py --config references/tasks/tabletop_pick_place.json --save-references
python3 scripts/record_overhead.py --config references/tasks/tabletop_pick_place.json --duration-sec 10.0- Configuration & Safety Invariants:
- All shared helper functions live in so101_common.py; never import internal monorepo-specific libraries inside plugin scripts or tests.
- Keep
max_step_deg <= 45.0(20.0–30.0default),max_speed_deg_per_sec <= 90.0(30.0default),joint_margin_deg >= 5.0(never blocking moves away from or back to arest_posethat sits inside the margin), andrest_tolerance_deg <= 5.0across bring-up motion configs. Stage 5 guarded touch-down (<= 12°/s) and hand-eye sweep (<= 15°/s) stay slow. - Leave STS3215 motor torque ON and print an explicit human warning if any exit path cannot verify return to
rest_pose.