Hardware setup and diagnostics for Hugging Face LeRobot — stable USB port mapping, motor-ID setup, a live motor monitor and calibration validation — inside a web workbench that also runs the full loop from teleop to policy evaluation.
Documentation · Direction · Contributing · Changelog · 한국어
LeLab is the official LeRobot GUI: calibrate, teleoperate, record, train and replay for the SO-101, with cloud training on HF Jobs. LeStudio does not try to replace it. What LeStudio adds is the part that comes before any of that works: getting hardware attached, identified and healthy.
| Need | Where to look |
|---|---|
| USB ports that change after every replug, arms you cannot tell apart | LeStudio Mapping — udev symlinks, Identify Arm wizard |
| Writing motor IDs one servo at a time | LeStudio Motor Setup wizard |
| Is this joint overloaded, in collision, or drawing too much current? | LeStudio Motor Monitor |
| Calibration file looks wrong | LeStudio Calibration — range / offset validation |
| Cameras: which is which, what FPS, what bus load | LeStudio Camera Setup |
| Record → train → evaluate on an SO-101 with one install | LeLab, or the LeStudio workbench pages |
| Live plots during teleop / eval | Foxglove (built into LeRobot 0.6.0) |
| Episode quality scoring at scale | LeRobot dataset visualizer, score_lerobot_episodes |
The hardware layer is a standalone package, lerobot-checkup, that the workbench depends on. Read Direction for the reasoning, the plan and what stays.
| Motor Setup | Status |
|---|---|
 |
 |
| Camera Setup | Dataset |
|---|---|
 |
 |
| Teleop | Train |
|---|---|
 |
 |
- Mapping: Camera and arm udev rules with stable symlinks, and an Identify Arm wizard (unplug, replug, assign).
- Motor Setup wizard: Write motor IDs one servo at a time with per-motor progress, error recovery and retry.
- Motor Monitor: Live position / load / current per motor, collision detection and clear, freewheel, E-Stop.
- Calibration: Run calibration, manage files, and validate ranges / homing offsets with errors and warnings.
- Camera Setup: Per-camera preview (MJPEG and snapshot), role assignment, FPS / bandwidth / bus utilization.
- Status Dashboard: Devices with mapped / unmapped state, prerequisites with a fix path, CPU/RAM/Disk/GPU.
- Preflight Checks: Validate devices, calibration, cameras, and CUDA before launch.
- Workbench Layout: Sidebar-driven workflow from hardware setup to training and evaluation.
- Global Console Drawer: Unified stdout/stderr stream, process input routing, and log copy actions.
- Layout: Desktop sidebar and a tablet drawer. Mobile widths are not supported.
- Config Profiles: Save, load, import, export, and delete working configurations.
- Session History: Track run-related events across recording, training, and evaluation flows.
- Design system: Semantic color tokens for both themes, enforced by a custom ESLint rule and a CI audit (see
packages/lestudio/frontend/DESIGN_GUIDE.md).
- Teleop: Multi-camera teleoperation with preflight checks and live SHM-shared camera feeds.
- Record: Episode recording with browser-side episode control, resume support, and preflight checks.
- Dataset: Local dataset listing, detail, delete, and quality checks.
- Episode Replayer: Multi-camera synchronized playback with timeline scrubbing.
- Episode Curation: Per-episode delete, tag, and filter.
- Hub Search & Download: Search and download datasets directly from Hugging Face Hub.
- Hub Push: Push local datasets with tracked job progress.
- Train: LeRobot training with CUDA preflight, real-time loss/LR chart, ETA tracking, and hyperparameter presets.
- Colab Integration: One-click Google Colab workflow for GPU-less environments.
- Dependency Remediation: Guided install flows for PyTorch and related training dependencies.
- Checkpoint Browser: Scan local checkpoints and auto-link to Eval.
- Eval: Policy evaluation with live process output and per-episode result tracking.
- Runtime Status: Shared WebSocket status, process stop controls, and orphan-process recovery signals.
- System Monitoring: GPU and system resource visibility from the UI.
- Error Translation: Common raw process failures converted into operator-readable guidance.
- Desktop Notifications: Browser notifications on process completion or error.
- Dark/Light Theme: CSS variable-based theme toggle.
- Python 3.12 recommended (3.10 / 3.11 work with lerobot 0.4.x)
- Linux (for
udevrules and/dev/video*access) lerobotis installed automatically as a dependency (see Installation)
- udev apply: one-click install works with either passwordless
sudoor a desktop Polkit auth prompt (pkexec). In headless/SSH environments without those, LeStudio provides manual commands. - Hub push / download:
huggingface-cli loginand a valid token are required. - GPU monitoring / CUDA preflight: CUDA environment and
nvidia-smirequired for full Train diagnostics.
LeStudio depends on upstream lerobot as a normal pip dependency (lerobot>=0.4.4,<0.7). Python 3.12 resolves the current 0.6.x line; Python 3.10 / 3.11 resolve 0.4.x, the last line that supports them.
Install from source:
git clone https://github.com/TheMomentLab/lestudio.git
cd lestudio
# one-time (if needed): conda create -n lerobot python=3.12 -y
conda activate lerobot
make installmake install installs lerobot-checkup and lestudio in editable mode and pulls lerobot with the dataset, training and Feetech / Dynamixel motor extras. If you need a specific torch build (CUDA version, CPU-only), install torch first and pip keeps it.
Upgrading from the submodule-based checkout: delete the old
lerobot/directory and re-runmake install. The fork is no longer used.
lestudioThe server starts at http://localhost:7860.
To open a browser automatically on desktop sessions, pass --browser (lestudio --browser or lestudio serve --browser). SSH or headless environments still skip browser opening.
usage: lestudio [-h] {serve,install-udev} ...
subcommands:
serve Start the LeStudio web server (default when no subcommand given)
install-udev Install udev rules via sudo (CLI alternative to the web UI)
lestudio serve:
--port PORT Server port (default: 7860)
--host HOST Server host (default: 127.0.0.1)
--lerobot-path PATH Path to lerobot source (auto-detected if installed)
--config-dir DIR Config directory (default: ~/.config/lestudio)
--rules-path PATH udev rules file (default: /etc/udev/rules.d/99-lerobot.rules)
--browser Open a browser automatically on startup
--no-browser Deprecated no-op; browser is not opened unless --browser is passed
--headless Alias for --no-browser
Flags can be passed without explicitly typing serve — lestudio --port 8080 works the same as lestudio serve --port 8080.
The hardware layer ships as its own package with a command line, installed together with LeStudio:
lerobot-checkup # hardware-only web UI (Status, Motor Setup, Camera Setup) on :7861
lerobot-checkup report # Markdown summary to paste into an issue
lerobot-checkup ports # serial ports that look like arms
lerobot-checkup cameras # cameras and the USB bus each one shares
lerobot-checkup motors --port /dev/ttyACM0 --ids 1-6
lerobot-checkup calibration # validate every calibration file in the LeRobot cache
lerobot-checkup udev status # are the stable /dev symlinks in place?Add --json to any command for machine-readable output. Details: packages/lerobot-checkup/README.md.
- Default bind is local-only:
127.0.0.1. - To expose on LAN, use:
lestudio serve --host 0.0.0.0. - When the UI is opened from another machine, use the header
Remotebadge to save the LeStudio session token for that server. A prompt still appears as a fallback on the first state-changing action if no token is stored yet. - Default CORS allows localhost origins only (
localhost/127.0.0.1).
You can override CORS with environment variables:
# Comma-separated explicit allowlist
export LESTUDIO_CORS_ORIGINS="http://localhost:7860,https://studio.example.com"
# Optional regex override (used when explicit origins are not set)
export LESTUDIO_CORS_ORIGIN_REGEX='^https://(localhost|127\\.0\\.0\\.1)(:\\d+)?$'For development compatibility only, LESTUDIO_CORS_ORIGINS="*" is supported but not recommended for shared networks.
packages/
├── lerobot-checkup/ # Hardware layer library: device discovery, udev/type/path policy, motor & calibration bridges
└── lestudio/ # Workbench: FastAPI backend (src/lestudio) + React frontend (frontend/)
tests/ # Backend tests for both packages
lestudio depends on lerobot-checkup; the dependency never points the other way. See docs_public/direction.md for why the repository is split this way.
conda activate lerobotFor contributor verification commands (ruff, mypy, pytest coverage helpers), install dev extras once:
make devBackend (with auto-reload on file changes):
lestudio serve --reloadBackend checks:
python -m ruff check packages
python -m mypy packages/lerobot-checkup/src/lerobot_checkup packages/lestudio/src/lestudio --ignore-missing-imports
python -m compileall -q packages/lerobot-checkup/src/lerobot_checkup packages/lestudio/src/lestudio
make testmake test scopes pytest to tests/ and sets PYTEST_DISABLE_PLUGIN_AUTOLOAD=1, which avoids unrelated plugins from the ambient environment. The backend CI job runs the same ruff, mypy, compileall, and pytest sequence before merge.
Frontend checks:
cd packages/lestudio/frontend
npm ci
npm run lint
npm test -- --run
npm run test:e2e
npm run buildnpm run build emits the frontend bundle to packages/lestudio/src/lestudio/static/, which FastAPI serves directly.
CI runs these checks automatically on every push: .github/workflows/ci.yml.
See CONTRIBUTING.md for architecture overview, PR guidelines, and the LeRobot import boundary rules.
Hardware smoke checks (real devices only, opt-in):
make test-hwWhen a pull request changes user-visible capabilities or top-level product messaging, update docs_public/feature-spec.md, README.md, and README.ko.md as part of the same change.
LeStudio began in February 2026 when LeRobot had no GUI. Hugging Face has since made LeLab the official one. The repository is therefore being reorganised around what remains open — hardware setup and diagnostics — as a standalone package, with the workbench kept on top of it. The full reasoning, the survey it rests on, and the step-by-step plan are in docs_public/direction.md.
- Status — Confirm cameras and arms are visible and process status is healthy.
- Motor Setup — Map devices to stable symlinks (udev rules), identify arms, run motor setup, and calibrate.
- Camera Setup — Verify camera streams and USB bandwidth.
- Teleop — Validate motion and camera feeds with preflight checks.
- Record — Capture episodes for your target task.
- Dataset — Inspect episodes, curate data, and push to Hugging Face Hub.
- Train — Start training and monitor loss/metrics in real time.
- Eval — Run policy evaluation to close the loop.
Apache 2.0 — see LICENSE.