Quickstart

End-to-end demo on a YAM bimanual rig and MolmoAct2-BimanualYAM. Same shape works for other YAML configs.

1. Install

pip install 'reflex-sdk[webrtc]'

The [webrtc] extra adds aiortc, av, msgpack, numpy, Pillow — the deps needed for the WebRTC inference transport. Plain pip install reflex-sdk gets you the CLI but no target.kind: webrtc support.

curl -LsSf https://astral.sh/uv/install.sh | sh
uv venv ~/reflex --python 3.11
source ~/reflex/bin/activate
uv pip install 'reflex-sdk[webrtc]'

For a YAM bimanual arm you also need the i2rt driver:

git clone https://github.com/i2rt-robotics/i2rt.git ~/i2rt
cd ~/i2rt && uv pip install -e .
sudo ip link set can0 up type can bitrate 1000000
sudo ip link set can1 up type can bitrate 1000000

See Installation for non-YAM hardware and macOS / Linux specifics.

2. Log in

reflex login

This opens a browser to app.tryreflex.ai/cli/authorize?requestId=... — approve the request and the CLI writes ~/.config/reflex/credentials.json with your API key. On a headless / SSH machine the link is Cmd+Clickable; copy it into your local browser if your terminal doesn't render OSC 8 hyperlinks.

reflex whoami
# Logged in to Reflex
#
# Organization: Acme Robotics
# Organization ID: org_...
# API key: laptop
# Dashboard: https://app.tryreflex.ai

3. Drop a YAML config

robot.yaml — bimanual YAM with three SHM-broadcast cameras pointed at the hosted MolmoAct2 worker:

mode: apply_actions
max_steps: 0
control_period_s: 0.2
stop_on_error: true
pipeline_inference: false

target:
  kind: webrtc
  base_model: molmoact2-bimanualyam
  connect_timeout_s: 180          # allow for cold start
  state_field: "state"
  instruction_field: "prompt"
  img_size: 256

hardware:
  kind: yam_bimanual
  config:
    left:  { channel: can1, gripper: linear_4310 }
    right: { channel: can0, gripper: linear_4310 }
    hz: 25
    instruction: "pack the container and close the box"
    home_duration_s: 4.0

cameras:
  top:   { kind: shm, name: top }
  left:  { kind: shm, name: left }
  right: { kind: shm, name: right }

Full field reference: Config reference.

4. Run

reflex connect --config robot.yaml

What you'll see:

[reflex] starting session: mode=apply_actions cameras=[top, left, right]
[webrtc-client] worker waking — holding for ready signal...
[webrtc-client] datachannel open
[webrtc-client] connected  session=sess_...
[reflex] hardware ready: connector=YamBimanualConnector ...
[reflex] step  0 │ apply ok
[reflex] step  1 │ apply ok
...

The first connect on a cold worker can take up to a couple of minutes while the model loads; reconnecting shortly after a session is fast — you're routed back to a warm worker.

Ctrl-C any time. Arms home automatically; a second Ctrl-C escalates to a force-exit.

For driver-level logs (i2rt motor power-on, aioice ICE candidates, per-step state/action vectors), add --verbose / -v.

What just happened

1. reflex connect parsed your yaml.
2. The SDK authorized a session and was routed to a ready worker (reusing a warm one if you'd recently connected).
3. The WebRTC transport waited for the worker to report ready, then opened a DataChannel.
4. After inference was confirmed reachable, the YAM connector powered the motors, started gravity-comp control loops, and calibrated the grippers.
5. The cameras attached to their /dev/shm segments.
6. The runner loop: read observation (state + 3 camera frames) → send over DataChannel → receive (30, 14) action chunk → apply the first N steps to the arms at 25 Hz → repeat.

Next