Skip to content

Reachy Mini Wireless · CM4 · daemon 1.10 · Strands

A robot that moves when it talks.

TINY is a Reachy Mini with a Strands agent inside. Four microphones that know where you are, a camera that finds your face, a six-axis head, two antennas and 81 recorded emotions — and a model that fires them while it speaks, not after.

Run it yourself The 27 tools Watch it look around ↓

TINY, a Reachy Mini: a cream rounded body on a dark printed foot, a flat cream head with two dark camera eyes, two thin black antennas — it turns its head toward you as you scroll
rendered from the twin's own meshes · curious1 t 0.0 shead yaw 5°roll 4°antennas 5 / 13° motor command yaw_body 0° stewart 0 0 0 0 0 0° antennas 0 / 0°
TINY mid-move: head turned 40° to its left, one antenna raised straight up, the other lying nearly flat
curious1 · t 2.0 s · head yaw 40° roll −6° · antennas 5 / 77° · body 8° pollen-robotics/reachy-mini-emotions-library · replayed through the daemon's IK + MuJoCo

81 recorded emotions · pick one

It looks around like it means it.

“You look around. Use this when you’re in a conversation with several people and want to glance at everyone.”

— the move’s own description, curious1
  1. 2.0 s head yaw 40°, one antenna snaps up to 77°
  2. 4.5 s yaw −37°, roll 14° — the other antenna drops to −85°
  3. 7.0 s the whole body turns 61°, head 45°
  4. 10.0 s yaw −45°, roll 16°, then…
  5. 11.8 s home. reachy_expression can fire any of the 81.

one realtime session · tools fire during speech

It moves while it speaks. Not after.

The voice persona is one Strands BidiAgent: microphone in, speech out, tool calls inside the same response. The gesture lands on the word, face-tracking lets go of the head only while it talks, and you can interrupt it.

  1. tiny.py:266 · the voice rules- Move WHILE you speak — gestures are simultaneous, never before/after.
  2. tools/reachy_expression.py:89Call it alongside speaking, just like neon fires arm gestures.
  3. tools/head_tracking.py:114 · SpeakingHandoffwhile the model streams speech audio → tracking hold "speaking" (weight 0.0 …); on response complete / interruption → release (weight 1.0) after a short playback tail.
  4. tools/xmos_audio.py:9 · XVF3800PP_NLATTENONOFF=0 suppression 17 dB → 4 dB; barge-in confirmed by the owner.
schematic of one turn — positions are not to scale; the tail is SPEAK_TAIL_S = 0.8 (tools/head_tracking.py:28), mic 16 kHz resampled to the model (tiny.py:451)
reachy_say("Hi, I'm TINY!") the offline voice · Piper en_US-lessac-medium on the CM4 (tiny-tts.service) · what the text personas sound like; the live call speaks with the realtime model · 1.3 soffline Piper voice · en_US-lessac-medium · 1.3 s
the same twin, one MuJoCo body group per layer · twin.xml · MuJoCo 3.13.0 · 41 meshes · 161 geoms · 9 actuators

18 bodies · 9 motors · one twin.xml

Nine motors. One face.

The dashboard's digital twin is the robot's own MuJoCo model, so this is not an illustration — it is the geometry the daemon solves against, split by body group.

  1. antennas two hinges, right_antenna · left_antenna — the fastest tell of a mood (curious1 throws one to 77° and the other to −85°)
  2. head body xl_330 with the camera_optical site — the face-finding camera; six degrees of freedom: roll, pitch, yaw and a little x y z
  3. Stewart platform six stewart_1…6 hinges on six rods with passive_1…6 ball joints — the head has no neck motor; six arms hold it up and tilt it together
  4. body yaw_body ±160° on the 3D-printed foot; inside the shell: a Raspberry Pi CM4, four microphones and the speaker

27 tools · 15 modules · @tool functions in tools/

Everything it can do is a function the model can read.

Every ability is a Python @tool with a docstring — the same text the model sees is what the reference prints at build time. Angles are clamped before they reach the daemon.

the owner's private cockpit · passkey gate · not linked here

The owner sees all of it.

A FastAPI dashboard runs on the CM4 itself behind a Cloudflare tunnel: the live camera, the same MuJoCo twin you just scrolled, every transcript and tool call from all personas, and the tracking controls. It is passkey-gated and not part of this page.

The cockpit on a desktop: camera view, transcript feed and controls.
desktop · camera, feed, controls
The cockpit showing the MuJoCo digital twin mirroring the robot's pose.
desktop · the twin, live
The cockpit on a phone.
phone

run it yourself · robot or MuJoCo sim

Six lines to a robot that talks back.

Bring a Reachy Mini and a Bedrock or OpenAI key — or no robot at all: make sim spawns a headless MuJoCo daemon.

Quickstart Showcase GitHub

git clone https://github.com/cagataycali/tiny-the-reachy.git
cd tiny-the-reachy
cp .env.example .env && $EDITOR .env   # your keys
make venv                              # .venv + deps
make run                               # REPL → the daemon
make sim                               # …or no robot at all