Skip to main content
IntroductionBuild AgentsAgent FrontendsTelephonyRoboticsWebRTC TransportManage & DeployReferences
XYouTubeSlack

Remote inference

Control your robot with models running in the cloud.

On this page

OverviewCommon patternsNext steps

Overview

For robots driven by end-to-end policies, LiveKit SDKs and LiveKit Portal let you run the policy in the cloud instead of onboard. Video streams and joint positions are sent to the model running in the cloud, and the resulting actions are streamed back to the robot for execution.

Running inference in the cloud lets you deploy models larger than a robot can run locally, update policies without reflashing hardware, and share compute across a fleet.

Common patterns

A remote inference loop runs in three stages:

  • Sensor input: Camera frames and joint or sensor state stream from the robot to the cloud.
  • Cloud policy: A model too large to run onboard computes the next action.
  • Action execution: Commands stream back to the robot fast enough to close the control loop.

Next steps

For implementation details, follow the inference guide:

Portal

Run inference in the cloud on realtime streams from robots connected through Portal.

Low-latency media

Tune video and data transport for latency-sensitive control loops.