Introduction

Reflex Connect streams observations from your robot to a hosted vision-language-action model and streams action chunks back, in real time. One YAML config, one command:

reflex connect --config robot.yaml

Under the hood that's a WebRTC DataChannel to a hosted GPU worker — UDP-based, so no TCP head-of-line blocking and low, consistent round-trip latency. The SDK owns the robot side: reads state and cameras, sends observations, applies returned action chunks at your configured control rate. You write zero networking, zero serialization, zero retry logic.

What you get

• Hosted GPU inference with fast warm-start: reconnect shortly after a session and you're routed back to a ready worker.
• Multi-region inference — the SDK automatically connects you to a low-latency region.
• First-class hardware support: YAM bimanual (via i2rt) and any V4L2 / RealSense camera. Shared-memory camera ingestion for zero-copy 3-camera 256×256 streaming.
• Safe defaults: motors never come live until the inference connection is confirmed. A Ctrl-C during connect aborts cleanly without powering hardware.

Who it's for

Roboticists deploying VLA policies on real arms. The connect path skips everything between "I have a robot and a hosted policy" and "the robot is moving" — no observation pipelines, no msgpack encoding, no DataChannel babysitting.

If you also need training jobs, datasets, or deployment provisioning, those live in the dashboard at app.tryreflex.ai. This docs site focuses on the inference connect surface.