A privacy-first home companion robot you build with your family.
Voice assistant. Object memory. Face recognition. Zero internet. All local.
What is Scout? • Build Phases • Quick Start • Privacy • Architecture • Contributing
Scout is an open-source home companion robot you build with your family. It patrols your house, remembers where objects are, greets family members by name, and answers questions - all without ever connecting to the internet.
"Scout, where are my keys?" "I last saw your keys on the kitchen counter about 2 hours ago."
The brain is an Arduino VENTUNO Q - a dual-processor board pairing a Qualcomm Dragonwing IQ-8275 (40 TOPS NPU, 16GB RAM) with an STM32H5 real-time controller. One board handles voice recognition, object detection, navigation, and face recognition. All locally.
Privacy is not a feature - it is the architecture. Scout has no Wi-Fi antenna pointed at the internet, no cloud account, no telemetry. Every byte of data stays on the robot. Face embeddings are encrypted at rest. Audio is never saved. You can audit the entire system with a single script.
Scout is designed to be built incrementally. Each phase produces a working robot. Start with Phase 1 and add capabilities over time.
Microphone, speaker, and a local LLM. One afternoon to build.
What can Scout do?
- Responds to "Hey Scout" wake word
- Answers questions using Llama 3.2 1B running on the NPU
- Tells jokes, sets timers, reads recipes
- Configurable voice and personality
- Runs from a desk or shelf - no wheels needed
Add a camera. Scout detects and identifies household objects.
What can Scout do?
- Detects objects via SmolVLM-256M on the NPU at 2 Hz
- "Scout, what do you see?" - describes objects in view
- Multi-object tracking with stable IDs across frames
- Optional ESP32-S3-CAMs for fixed room monitoring
Wheels, motors, LIDAR, and autonomous navigation.
What can Scout do?
- Maps your house using RTAB-Map SLAM
- Navigates between rooms via Nav2
- Runs patrol routes on a schedule
- 4-layer obstacle avoidance (LIDAR + ToF + cliff + bump)
- Hardware e-stop button kills all motor power instantly
- STM32H5 watchdog stops motors if Linux crashes
Software-only. Scout remembers where it saw household objects.
What can Scout do?
- "Scout, where are my keys?" - tells you the room, zone, and when
- SQLite + FTS5 spatial database with confidence decay
- Portable items decay fast (4h half-life), furniture decays slow (1 week)
- Natural language aliases: "my phone" maps to "phone"
- Configurable retention (default 30 days)
Software-only. Face recognition with encrypted embeddings.
What can Scout do?
- "Good morning, Sarah!" - greets family by name
- Time-of-day aware greetings with per-person cooldown
- Consent-required enrollment (5 seconds in front of camera)
- 128-d face embeddings encrypted at rest (AES-256-GCM)
- No face images ever saved
Indoor micro drone companion for elevated views.
What can Scout do?
- Small indoor drone streams video back to Scout for VLM inference
- Checks high shelves, tops of cabinets, above furniture
- Prop guards required, strict flight boundaries
- This phase is a design outline - contributions welcome
Total cost across all phases: ~$1,128 (or ~$362 for voice-only Phase 1)
git clone https://github.com/m4cd4r4/home-scout.git
cd home-scout
docker compose up simgit clone https://github.com/m4cd4r4/home-scout.git
cd home-scout
./scripts/setup-ventuno.sh
source /opt/ros/jazzy/setup.bash
source ros2_ws/install/setup.bash
ros2 launch scout_bringup scout_voice_only.launch.pySay "Hey Scout" and ask a question. See the Phase 1 Build Guide for the full walkthrough.
| Principle | Implementation |
|---|---|
| No internet | Scout creates its own WiFi AP (ScoutNet). No default gateway. No DNS. |
| All AI local | Whisper, Piper TTS, Llama 3.2 1B, SmolVLM, ArcFace - all on the 40 TOPS NPU |
| No telemetry | Zero analytics, zero crash reports, zero usage data |
| Encrypted faces | 128-d embeddings encrypted with AES-256-GCM. Key from user passphrase via Argon2id |
| Audio not saved | Processed in circular buffer, discarded after transcription |
| Hardware mic mute | Physical slide switch disconnects I2S clock line |
| Camera LED | Wired to camera power. If the LED is off, the camera is off. Hardware guarantee. |
| Auditable | scripts/verify-privacy.sh checks network isolation, DNS, traffic, listening services |
+----------------------------------------------------------+
| VENTUNO Q (Ubuntu 24.04 + ROS 2 Jazzy) |
| |
| +----------+ +----------+ +---------+ +------------+ |
| | voice | | vision | | nav | | memory | |
| | Phase 1 | | Phase 2 | | Phase 3 | | Phase 4 | |
| +----------+ +----------+ +---------+ +------------+ |
| +----------+ +--------------------------------------+ |
| | faces | | hardware_bridge | |
| | Phase 5 | +------------------+-------------------+ |
| +----------+ | CAN-FD |
+-----------------------------------+----------------------+
| STM32H5 (Zephyr RTOS) |
| Motor PWM | Encoders | PID | Safety Watchdog |
+----------------------------------------------------------+
| Model | Purpose | Size |
|---|---|---|
| openWakeWord | "Hey Scout" detection | 2 MB |
| Whisper-Small-En | Speech recognition | 500 MB |
| Piper TTS (amy) | Speech synthesis | 100 MB |
| Llama-3.2-1B-Instruct | Conversation | 1.5 GB |
| SmolVLM-256M | Object detection (VLM) | 1.5 GB |
| SCRFD-2.5G | Face detection | 10 MB |
| MobileFaceNet | Face embedding | 10 MB |
Total: ~3.6 GB disk, ~5.4 GB RAM. Leaves 10.6 GB free on the 16 GB board.
See ARCHITECTURE.md for full system design, NPU scheduling, object memory schema, and network topology.
%%{init: {'theme': 'dark', 'themeVariables': {'primaryColor': '#0f172a', 'primaryTextColor': '#e2e8f0', 'primaryBorderColor': '#3b82f6', 'lineColor': '#64748b'}}}%%
graph TB
classDef hw fill:#1c1400,stroke:#f97316,color:#fdba74,font-weight:bold
classDef sense fill:#0f1f3d,stroke:#3b82f6,color:#93c5fd
classDef npu fill:#160d2e,stroke:#8b5cf6,color:#c4b5fd,font-weight:bold
classDef ros fill:#0a2520,stroke:#10b981,color:#6ee7b7
classDef stm fill:#1c0a1c,stroke:#ec4899,color:#f9a8d4
subgraph SENSORS["Sensors"]
Mic["🎤 Dual INMP441\nI2S mics"]:::sense
Cam["📷 Arducam IMX477\nHQ MIPI-CSI"]:::sense
LIDAR["⭕ RPLIDAR C1\n360° SLAM"]:::sense
ToF["📡 4× VL53L1X\nToF proximity"]:::sense
IMU["🧭 BNO055 IMU"]:::sense
end
subgraph VENTUNO["VENTUNO Q — 40 TOPS NPU · 16 GB RAM (Ubuntu 24.04 + ROS 2 Jazzy)"]
WW["openWakeWord\nHey Scout"]:::npu
ASR["Whisper-Small\nSpeech → Text"]:::npu
LLM["Llama 3.2 1B\nConversation"]:::npu
VLM["SmolVLM-256M\nObject Detection"]:::npu
Face["SCRFD + MobileFaceNet\nFace Recognition"]:::npu
Nav["ROS 2 Nav2\nSLAM + Path Plan"]:::ros
Mem["Object Memory\nSQLite encrypted"]:::ros
TTS["Piper TTS\nText → Speech"]:::npu
end
subgraph STM["STM32H5 — Zephyr RTOS"]
Motor["Motor PWM · PID"]:::stm
Safety["Safety Watchdog\nhardware e-stop"]:::stm
end
Mic --> WW --> ASR --> LLM --> TTS
Cam --> VLM --> Mem
Cam --> Face --> Mem
LIDAR & ToF & IMU --> Nav
LLM --> Nav
Nav -->|"CAN-FD"| Motor
Safety -.->|"override"| Motor
| Part | Model | Cost |
|---|---|---|
| Main board | Arduino VENTUNO Q | ~$300 |
| Storage | Samsung 970 EVO Plus 250GB NVMe | ~$40 |
| Audio | 2x INMP441 mics + MAX98357A amp + speaker | ~$17 |
| Camera | Arducam IMX477 HQ MIPI-CSI | ~$60 |
| Motors | 4x Pololu 37D 50:1 with encoders | ~$100 |
| Chassis | 3D printed or off-the-shelf 4WD kit | ~$35 |
| LIDAR | SLAMTEC RPLIDAR C1 | ~$100 |
| Sensors | BNO055 IMU + 4x VL53L1X ToF + 2x TCRT5000 cliff | ~$90 |
| Power | 4S LiPo + BMS + buck converter | ~$97 |
| Connectors | JST-XH, cables, standoffs, fuse | ~$70 |
Full BOM with 3 price tiers and alternatives: BOM.md
home-scout/
ros2_ws/src/ # 8 ROS 2 packages (voice, vision, nav, memory, faces, ...)
firmware/ # STM32H5 motor control + ESP32-CAM streaming
hardware/ # 3D printable chassis, wiring schematics, mounts
config/ # Personalities, patrol routes, room maps
docs/ # Build guides (6 phases), reference, design docs
training/ # Wake word, object detection, face recognition guides
tests/ # Unit, integration, simulation, hardware tests
scripts/ # Setup, model download, privacy verification
docker/ # Dev and simulation environments
Contributions of all kinds are welcome - code, docs, hardware designs, 3D models, testing.
git clone https://github.com/m4cd4r4/home-scout.git
cd home-scout
docker compose up dev- Write tests using the Gazebo simulation
- Improve documentation and build guides
- Design 3D-printable chassis variants
- Optimize AI models for edge deployment
- Add language support (wake words + TTS voices)
- Review code for security and privacy
See CONTRIBUTING.md for the full guide.
Read SAFETY.md before building. Key points:
- E-stop is mandatory for Phase 3+. Cuts all motor power instantly.
- LiPo battery handling requires adult supervision. Never charge unattended.
- Phase 6 (drone) is adults-only. Indoor drones require careful tuning.
- Max speed software-limited to 0.3 m/s (enforced independently by STM32H5).
- STM32H5 watchdog: if Linux crashes, motors stop within 100ms.
MIT. Build it. Modify it. Share it. Teach your kids how robots work.
Scout never phones home. That is the point.