Skip to content
MouseMouse
Menu

Inference

Deploy policies to robots, or serve them from a GPU. A person approves every run on a real robot.

Deploy a policy to your robotHow approvals work

Two paths, by the policy’s speed

Where a policy runs depends on its speed. A policy answers in milliseconds, so it runs on the robot itself. A policy is too large for the robot’s computer and answers in chunks of actions, so a GPU serves it. Both reach a robot the same way: a request, a person's approval, a signed trail, and a Stop button.

Deployed to the robot, or served from a GPU

The two paths a policy takes to a robot, row by row
WhatDeployed to the robotA reflex policyServed from a GPU endpointA skill policy
The policySmall and fast, in one ONNX file, at 50 Hz: walking, balance, holding a pose.Larger, reads camera images and a task, at 1 to 10 Hz: a vision-language-action policy such as SmolVLA.
Where it runsOn the robot's own computer, from a file it downloads.On a GPU the robot calls over the network: your own, or Lucen cloud.
How it startsSomeone asks to run it on one robot for a set number of minutes. A signed-in person approves. The robot checks the approval's signature itself, downloads the exact file by its checksum, and runs it.Start an endpoint from the policy's Serve tab, or serve it from your own GPU. The robot opens a session with an approval a person signed. The session is bound to that robot, that policy's checksum and that endpoint, and metered.
Who approvesA signed-in person. An API key cannot, whatever its scope, so your AI can ask and then wait.The same person, the same way. A key cannot point a robot at an endpoint.
How it stopsStop now on the approval's page, or STOP on My robots. The window's end. Ten seconds without the hub: the robot holds, then stops on its own.The same stop. An endpoint also ends when a person stops it, or after ten idle minutes on Lucen cloud. If the link is lost, the robot holds still at once, then stops.
What it costsFree. The robot runs it.Free on your own GPU. On Lucen cloud, the GPU's hourly rate, metered per session.

Start here