Developer platform · physical AI

The runtime network for live intelligence.

See Act People in the loop

Purpose-built for video, audio and sensor streams. Your models, devices and people in one session.

Get early access See how it works
rgb.live / sessions / warehouse-east LIVE
depth · detector · agent 30 fps
Works with

Interactive, low latency, ready to run.

Full-duplex sessions, latency you can act on, and a catalog of models optimized for live video.

01 · FULL DUPLEX

Interactive, in both directions

Video, audio, results and control flow both ways in one session. Change a parameter, prompt the model, hand control to a person or take it back, all live, without reconnecting. Sessions built to run for a month, not a meeting.

02 · LOW LATENCY

Fast enough to act on

Overlays, results and control come back inside a frame budget, not after a round trip to an inference API. The media and the model share a GPU, so there is no hop between seeing and deciding.

03 · CURATED CATALOG

Optimized models, ready to run

Vision-language agents, detection, pose, depth, segmentation, transcription and world models, tuned for live video. Pick one and have a moving picture in minutes, or bring your own weights.

How a session is wired

Built for developers

One SDK call to join. One session to run everything.

Create a session from your backend, invite your sources and your model, and subscribe to what comes back. No media pipeline to build, no GPU fleet to schedule, no backend for the first moving picture.

  1. 1
    Create a session

    From your backend, with the model you want in it: one from the catalog, or your own weights.

  2. 2
    Invite the participants

    Cameras over RTSP, drones over WHIP, robots over ROS 2, phones and browsers over the SDK. People, when they need to see.

  3. 3
    Subscribe to what comes back

    Processed video, structured results per frame, alerts and control, in both directions, live.

SDKs
  • JavaScript · browser
  • Node.js · server
  • Python
  • Go
  • C++
  • Flutter
Transports and protocols
  • WebRTC
  • WHIP in
  • WHEP out
  • RTSP bridge
  • ROS 2
  • WebSocket
  • WebTransport
In the catalog
  • Qwen3-VL agents
  • Qwen3
  • Gemma 4
  • YOLO11 detection
  • YOLO11 pose
  • Depth Anything
  • SAM 3 segmentation
  • Whisper
  • Cosmos 3
  • LingBot-VA

Three things a session does.

See, act, learn. The three bars in the mark. One session, one clock and one SDK carry cameras, drones and robots, so what you build for fixed cameras today runs on a moving one tomorrow.

PHASE 1 · SEEEarly access
Cameras and a drone streaming into one session

Visual intelligence with people in the loopMany live sources, your model, a supervisor who can answer.

Every camera, drone and phone publishes into one session. Your model watches all of them and sends back overlays, depth and alerts. A supervisor joins from a browser, sees the alert beside the live view, and clears or escalates it.

Who
Camera integrators and ISVs, drone and public-safety operators, security operations centres.
Priced
Per stream per month, 24/7, your model included.
PHASE 2 · ACTEarly access
A robot arm and a rover controlled through a live session

Remote operation and controlControl streams out on the session that carries the video.

Control rides the same session as the video, with takeover and handback built in. Operate drones, arms and vehicles remotely while the model and the operator see the same frame.

Who
Robot OEMs, robot and drone fleets, machine makers.
Priced
Per device per month.
PHASE 3 · LEARNEarly access
A film strip of recorded frames with one episode highlighted

Recording, replay and episodesEvery session on one clock becomes training data.

Every session is recorded on one clock, replayable against a new model and exported as episodes. Every takeover becomes a label, captured as a by-product of running the session.

Who
Foundation-model labs, humanoid makers, data vendors.
Priced
Per recorded hour plus storage.

Bring your streams →

Also runs today, as demos Depth to 3DBackground blur and poseVoice agent and meetingsPhysics and world simulation

Every stream. Every person. One session.

Stop stitching cameras to inference APIs to dashboards to pagers. Put the model and your people in the same room as the stream.

100+
streams per model

One model, every stream.

Every source publishes into the session and your model subscribes to all of them. Add a camera by inviting it.

Live
people in the loop

Your people see what the model sees.

A supervisor joins from a browser, sees the alert beside the live view, and clears or escalates. Models triage. People decide.

Agentic
models that act

Models that do, not just detect.

A vision-language agent in the session watches, reasons, calls a tool, raises an alert or hands off to a person. Prompt it live, mid-stream.

<150 ms
glass to glass

Fast enough to act on.

No hop between seeing and deciding. Perception and control ride the same session, so an operator and a policy work from the same frame.

Get early access →

Design targets, p95: under 150 ms glass to glass on LAN or 5G, under 300 ms over the public internet, under 100 ms per frame in the model loop.

Early access · 2026

Your streams, one session. Running in minutes.

We are onboarding a small number of teams this quarter. Bring the cameras, drones or robots you already have; we run the session and you watch the number every week.

Your streams. Ten streams minimum, a hundred welcome. Pick a model from the catalog and go.
We run it. Our GPU nodes, or yours on site. A supervisor seat for the person who answers alerts today, from day one.
A number every week. Glass-to-glass latency and streams running, from the session itself, in writing.
Free for the first 30 days. Then per stream per month, with early-access pricing locked for the first year.
Built for
Integrators and ISVs with a hundred to ten thousand cameras and a model of their own, who want to run it on every stream without operating a GPU fleet.
Drone inspection and public-safety operators whose pilots and analysts watch live feeds and want a model in the loop.
Security operations centres and alarm-verification teams that pay people to look at events and want the model to triage first.
Robot and machine fleets that want perception and control on one session, with takeover and handback built in.
Open in email Send it to hello@rgb.live