Reflex: Fastest GPU Cloud for Robot Inference

SDK

pip install 'reflex-sdk[webrtc]'

Install	What you get
`pip install reflex-sdk`	CLI, `target.kind: edge` (local HTTP) and `target.kind: platform` (platform-routed) transports.
`pip install 'reflex-sdk[webrtc]'`	Above + `target.kind: webrtc` (DataChannel to hosted GPU workers). Adds `aiortc`, `av`, `msgpack`, `numpy`, `Pillow`.

Requires Python ≥ 3.9. Tested on 3.10–3.12.

Hardware drivers

The SDK ships connectors for:

Connector kind	Hardware	Driver dep
`subprocess`	Anything — script-based connector via stdin/stdout JSON	none
`yam_bimanual`	i2rt YAM bimanual arms over SocketCAN	`i2rt`

YAM bimanual

i2rt isn't on PyPI; clone + install editable. From a fresh Ubuntu 22.04+ machine:

sudo apt update && sudo apt install -y \
    build-essential python3-dev linux-headers-$(uname -r) \
    can-utils iproute2 git curl

# uv (Astral) is the recommended Python toolchain.
curl -LsSf https://astral.sh/uv/install.sh | sh

# i2rt — YAM connector + motor drivers.
git clone https://github.com/i2rt-robotics/i2rt.git ~/i2rt
cd ~/i2rt
uv venv --python 3.11 .venv
source .venv/bin/activate
uv pip install -e .

# Reflex SDK into the same venv.
uv pip install 'reflex-sdk[webrtc]'

# CAN bus — one-time per host. Adjust interfaces to match your canable2 setup.
sudo ip link set can0 up type can bitrate 1000000
sudo ip link set can1 up type can bitrate 1000000

# Auto-enable on boot:
cd ~/i2rt && sudo sh devices/install_devices.sh

Verify:

ip -br -c link show | grep can       # can0 UP, can1 UP
which reflex                          # ~/i2rt/.venv/bin/reflex
reflex whoami                         # exercises the CLI end-to-end

Cameras

Camera kind	Source	Notes
`v4l2` (alias: `webcam`)	`/dev/video*` USB cameras	`cv2.VideoCapture` under the hood
`realsense`	Intel RealSense via `pyrealsense2`	Install RealSense SDK + udev rules separately
`shm`	Pre-broadcast frames from `/dev/shm/reflex_cam_*`	Multi-consumer-safe; useful when several workers read the same camera

The SHM source is read-only — the broadcaster (not part of the Reflex SDK) publishes frames. Typical setup is a camera_broadcaster.py you author that grabs from RealSense / V4L2 and writes raw HxWxC uint8 into a memory-mapped segment per camera. See the YAM rig setup notes for a reference broadcaster.

macOS

i2rt is Linux-only (evdev is gated on sys_platform == 'linux', and the CAN-over-USB drivers are different on macOS). The Reflex SDK itself works on macOS — you can run reflex login, reflex whoami, hit edge workers, drive an SO-101 via the subprocess connector — but the yam_bimanual connector won't import.

ARM Linux (Jetson, etc.)

aiortc wheels are x86_64 only. On ARM, pip will try to compile PyAV from source, which needs ffmpeg dev headers:

sudo apt install -y libavcodec-dev libavformat-dev libavutil-dev \
    libswscale-dev libavfilter-dev libavdevice-dev libswresample-dev \
    libsrtp2-dev libssl-dev

Configuration

reflex resolves URLs in this order:

Slot	Backend	Platform
`--convex-url` / `--site-url` flag	✓	✓
`REFLEX_CONVEX_URL` env	✓	—
`REFLEX_PLATFORM_URL` env	—	✓
`~/.config/reflex/credentials.json` (set by `reflex login`)	✓	✓
Default	`https://api.tryreflex.ai`	`https://app.tryreflex.ai`

A fresh install does not need any env var set. Defaults cover the production deployment; flags / env override for staging / self-hosted.

Troubleshooting

Symptom	Likely cause
`API key is required`	Run `reflex login` first, or set `REFLEX_API_KEY`.
`Convex URL must start with http:// or https://.`	A `--convex-url` / `REFLEX_CONVEX_URL` override isn't a full URL. Use `https://...`.
`ModuleNotFoundError: No module named 'aiortc'`	Install with `[webrtc]` extra: `pip install 'reflex-sdk[webrtc]'`.
`RuntimeError: WebRTC signaling timeout` on first connect	Worker cold-start exceeded the budget. Bump `connect_timeout_s` to 180 in your yaml.
ICE never completes	UDP egress blocked. Try a different network; check firewall for outbound UDP.

For non-obvious failures, run with --verbose:

reflex connect --config robot.yaml --verbose

Installation

SDK