Lanturn 🔦 (Work in Progress)

A hackathon project connecting the Gemini Live API to an ESP32 Atoms3r-CAM device for voice + vision conversations on embedded hardware.

Overview

Lantern demonstrates real-time AI voice conversations with vision running on an ESP32 microcontroller. It uses:

• ESP32 Atoms3r‑CAM (mic, speaker, camera)
• Pipecat (voice/media orchestration)
• Gemini Live API (multimodal speech + vision)

Features

• ✅ Real-time voice conversations
• ✅ Real-time vision processing with camera (Work in Progress)
• ✅ Gemini Live multimodal AI integration
• ✅ ESP32 hardware support
• ✅ WebRTC audio + vision streaming
• ✅ Automatic greeting on connection
• ✅ Google Search tool call
• ✅ AI can see and describe what the camera shows

Architecture

ESP32 Atoms3r-CAM  <---WiFi/WebRTC--->  Pipecat Server  <--->  Gemini Live API
(Camera + Mic + Speaker)              (Signaling + Bot + Processing)       (Multimodal LLM)

Data Flow

1. ESP32 captures audio from microphone and camera video
2. ESP_H264 converts the camera frames using Espressif’s lightweight H.264 software encoder to H.264 QVGA (320x240) at ~1 FPS and streams via a WebRTC video track; audio uses Opus
3. Pipecat handles WebRTC signaling (/api/offer) and communicates with server to server via Gemini Live API
4. Gemini generates voice responses based on what it hears AND sees (inference)
5. Audio streams back to the tiny devices speaker

Prerequisites

Hardware

• ESP32 Atoms3r-CAM (M5Stack) with GC0308 camera
• USB-C cable for programming/power
• WiFi network on the 2.4ghz band with AP Isolation Disabled

Software

• Python 3.13+
• uv (Python package manager)
• ESP-IDF toolchain (installed at ~/esp/esp-idf)

Quick Start

1. Install server deps

cd lanturn
uv sync

1. Configure API key (one of)

export GOOGLE_API_KEY=your_api_key_here
# or
cp env.example .env

1. Run the server (SmallWebRTC on :7860)

uv run python Lanturn_esp32_gemini_live_vision_bot.py -t webrtc --esp32 --host YOUR_COMPUTER_IP

1. Flash ESP32 firmware

cd esp32/pipecat-esp32/esp32-m5stack-atoms3r
source ~/esp/esp-idf/export.sh
idf.py set-target esp32s3
export WIFI_SSID="your_wifi_network"
export WIFI_PASSWORD="your_wifi_password"
export PIPECAT_SMALLWEBRTC_URL="http://YOUR_COMPUTER_IP:7860/api/offer"
idf.py build -p PORT flash

1. Monitor

idf.py monitor

Setup

1) Start the Pipecat server (Python)

Follow the Quick Start steps (1–3).

• The ESP32 will POST its WebRTC offer to: http://YOUR_COMPUTER_IP:7860/api/offer
• Ensure your firewall allows inbound connections to port 7860 on YOUR_COMPUTER_IP.

Notes:

• The server enables audio-in, audio-out, and video-in; video-out to ESP32 is disabled.

2) Build and flash the ESP32 firmware

Set these env vars (compiled in): WIFI_SSID, WIFI_PASSWORD, PIPECAT_SMALLWEBRTC_URL Don’t use quotations, just values. (see Quick Start step 4).

What to look for in the logs:

• WiFi connects and shows an acquired IP
• “Camera initialized successfully” (or a warning if not present)
• “Camera streaming enabled”
• Periodic logs like “Sent N H.264 frames”

Usage

1. Start server. 2) Power on ESP32. 3) Speak and point the camera. To test for Vision, Try:

• “What do you see?”
• “Describe what’s in front of you”
• “What color is this?”
• “Read the text you see”

*Be specific about the video because the model will often hallucinate if there’s no camera feed, and tell you otherwise.

Project Structure

lanturn/
├── Lanturn_esp32_gemini_live_vision_bot.py  # Vision-enabled bot (Pipecat + Gemini Live)
├── Lanturn_esp32_gemini_live_alt_bot.py     # Audio-only alt bot
├── main.py                                  # Placeholder entry point
├── env.example                              # Sample env file
├── pyproject.toml                           # Python dependencies
└── README.md                                # This file

Documentation

• pipecat.ai docs: https://docs.pipecat.ai
• Gemini Live: https://ai.google.dev/gemini-api/docs/multimodal-live

Technical Details

Vision Specifications

Camera: GC0308 0.3MP sensor

Resolution: QVGA (320x240)

Video Codec: H.264 Baseline

Frame Rate: ~1 FPS

Bitrate: ~200 kbps target

Transport: SmallWebRTC video track (H.264) + Opus audio

Dependencies (ESP-IDF component registry):

• espressif/esp32-camera (GC0308, RGB565 capture)
• espressif/esp_h264 (software H.264 encoder)

Build notes:

• ESP‑IDF 5.5.x, PSRAM required (8MB on AtomS3R‑CAM)
• Video task pinned to Core 1; audio on Core 0
• Camera configured RGB565 QVGA; converted to I420 for encoder
• RTP video timestamp set to 1 fps for accurate pacing

Pin map (GC0308) from M5 docs:

• CAM_SDA=G12, CAM_SCL=G9, VSYNC=G10, HREF=G14, XCLK=G21, PCLK=G40, Y9=G13, Y8=G11, Y7=G17, Y6=G4, Y5=G48, Y4=G46, Y3=G42, Y2=G3, POWER_N=G18

Performance notes:

• ESP32‑S3 software H.264 is viable at low FPS/QVGA; IDR every frame for robustness
• Use PSRAM for camera and encoder buffers
• Expect ~1–2s end-to-end visual latency

Known Issues

• ESP32 camera may not power on in some setups. We suspect an init/power sequencing issue or hardware power constraint. Contributions with stable init/power routines for AtomS3R‑CAM are welcome.

Development

Running Locally

uv run python Lanturn_esp32_gemini_live_vision_bot.py -t webrtc --esp32 --host YOUR_IP

Testing Without Hardware

uv run python Lanturn_esp32_gemini_live_vision_bot.py -t webrtc --host YOUR_IP
# Then open: http://YOUR_IP:7860

Monitoring ESP32

cd esp32/pipecat-esp32/esp32-m5stack-atoms3r
source ~/esp/esp-idf/export.sh
idf.py monitor

Credits

• Pipecat: pipecat.ai
• Pipecat-esp32: pipecat.ai
• Gemini Live: Google AI
• ESP-IDF: Espressif
• M5Stack: m5stack.com
• ESP32-Camera: esp32-camera library

Resources

• AtomS3R-CAM Documentation
• ESP H.264 use tips
• Gemini Live API
• Pipecat Documentation

License

BSD 2-Clause; see LICENSE.

Hackathon Info

• Project: Lanturn
• Goal: Voice + Vision AI on ESP32 embedded hardware
• Stack: ESP32 Atoms3R-CAM + Pipecat + Gemini Live
• Status: Functional prototype with voice (vision not working yet)
• Sponsor: Google and Pipecat hackathon

---

Made with ❤️ for the Google + Pipecat Hackathon