Ship robots,
not infrastructure.

The web control plane for LeRobot developers.
Connect your robot, collect data, train models, and deploy — all from one browser.

Web Serial / WebUSB LeRobotDataset v3 HuggingFace Hub GPU Training VLA Inference
armpilot.sh
# Before ArmPilot: 7 tools, 3 repos, zero visibility
$ ls ~/lerobot-projects/
cli-scripts/    docker-setup/    data-chaos/
$ python train.py --config custom.yaml
RuntimeError: CUDA out of memory.
 
# With ArmPilot: one URL, full workflow
→ Connect robot at app.armpilot.dev
→ Collect 50 episodes, push to Hub
→ Train π0 policy, deploy to hardware
The Workflow

One page. Full robotics pipeline.

01

Connect

Pair your SO-101, ALOHA, or any LeRobot-compatible arm via Web Serial or WebUSB. Stream live video directly in the browser.

02

Collect

Teleoperate and record structured episodes. Data saves in LeRobotDataset v3.0 format — parquet + synchronized video — and syncs to HuggingFace.

03

Train

Spin up GPU training jobs — GR00T, π0, diffusion policies — without managing cloud infrastructure. Monitor progress in real-time.

04

Deploy

Run inference and send actions back to your robot. Serve models, log performance, iterate — all from one dashboard.

Built for developers

Everything the LeRobot ecosystem needs, unified.

Browser-native robot control

No CLI, no local setup. Connect your robot over USB, stream camera feeds, and teleoperate directly from Chrome or Edge. Works on Linux, Mac, and Windows.

LeRobotDataset management

Visualize, edit, split, and version datasets. Episode browser, frame-level inspection, and one-click HuggingFace sync with automatic v3.0 parquet formatting.

Managed GPU training

Train VLAs without managing cloud configs. GR00T, π0, diffusion policies — submit a job, watch the loss curve, pull the checkpoint when it's done. Per-second billing.

Model-agnostic backend

GR00T, π0, π0.5, GEN-1, Gemini Robotics — future foundation models plug in as execution backends. ArmPilot owns the workflow layer regardless of which model runs underneath.

Real-time inference

Deploy trained policies to your hardware with low-latency inference. Log action trajectories, monitor drift, and compare policy versions side-by-side.

Open-source native

Built on LeRobot's Python stack. Your workflows are reproducible, exportable, and portable. No vendor lock-in — data and models live on HuggingFace Hub where you control them.

Supported Hardware

Your robot. Your rules.

ArmPilot supports any LeRobot-native hardware out of the box. Connect and start working — no firmware flashing, no custom drivers.

SO-101 / SO-100

Hugging Face's 3D-printable arm. The primary entry point for LeRobot developers worldwide.

ALOHA / ALOHA-2

Dual-arm manipulation platform. Bimanual teleoperation with synchronized leader-follower control.

StarAI Arm

Seeed Studio's 6+1 DoF arm with LeRobot plugin support. Integrated ZED camera depth perception.

Humanoids

HopeJR, xArm, and emerging humanoid platforms. ArmPilot scales with the ecosystem.

Don't see your robot? LeRobot's plugin system means ArmPilot supports any new hardware with a pip install.

Robotics developers should not spend weeks building infrastructure before making a robot move.

LeRobot is becoming the open-source standard for robot learning. ArmPilot makes it usable by everyone — from researchers publishing at CoRL to hobbyists building their first agent arm.