Getting Started

DeviceNexus is the operational layer that connects the tools you already use — teleoperation, cloud GPUs, simulation, OTA — into one pipeline: teach → train → validate → deploy → improve, across a fleet of robots, from a browser.

This page is the honest account of what you can do today and where the frontier is.

What's proven today

The record → train → publish a checkpoint loop is verified end-to-end through the Console UI on a real SO-101 arm. Delivering that checkpoint to a robot over the air and hot-swapping it is also verified. Running the policy on a real arm and measuring on-task success — and the fleet-wide “keeps getting better” loop — is the active frontier, not yet a finished hardware result. We would rather you know exactly where the line is.

Before you start

You will get the most out of a pilot if you have:

• A robot we can talk to — an SO-101 arm works out of the box; any ROS 2 robot works via the ros2 adapter; you can also start fully in simulation (MuJoCo) with no hardware.
• A Linux machine with a GPU for training — you bring the compute; we don’t resell it. (You can also start with recording and dataset work before any training.)
• An invite — access is currently invite-gated while we onboard design partners. If you don’t have one, email sales@devicenexus.ai.

The five-minute path

1. Sign up and sign in. Use your invite link to create an account and organization. You’ll land in the Console.

2. Enroll a device. Open the device onboarding wizard, pick a Device Profile (search + type filter), reveal its enrollment token, and run the one-line install command on your machine. The agent, connectivity, and capture components install automatically — the device type derives from the profile you chose.

3. Record a few episodes. In the Teach tab, drive the robot (physical leader arm, keyboard/gamepad, sim, or VLM) and record. Each episode is captured as an MCAP recording and converted to a training-ready dataset, with the data visualized inline.

4. Train a policy. From the Training page, point a job at a dataset and pick a model family (ACT, Diffusion, and SmolVLA are live-trained; Pi0 / GR00T N1.7 / MolmoAct2 are supported). Training runs on your GPU; loss curves and dataset previews show in the Console.

5. Publish and deploy. When a checkpoint is ready, deploy it. The platform delivers it to the robot over the air and stages it (idle, torque held) — running it is an explicit, separate operator action, and an e-stop halts a running policy.

What happens behind the scenes

The platform only ever sees observations (camera frames + joint states) and actions (joint commands). It does not need to know it’s an SO-101 or that it’s sorting blocks — that’s why the same pipeline works across robots and tasks. Recordings upload to the platform’s catalog; training writes checkpoints to object storage; robots resolve and download them via short-lived signed URLs. Your data, datasets, and trained models remain yours (see Security and the Privacy Policy).

Where to go next

• API Reference — the REST API for automating enrollment, datasets, and deployments.
• Security — how the platform is built and operated, and our honest compliance posture.
• Glossary & Learning Path — concepts from zero to running learned policies.
• Stuck? Email support@devicenexus.ai — a person reads it.

Honest limits

A simulation pass is a filter, not a hardware guarantee. The platform has not been load-tested beyond a handful of devices. Humanoid teleoperation and some model families are still maturing. We keep a running, honest list — ask us anything a pilot needs to know before you commit.