Reachy Mini Robot with NeMo Agent Toolkit Tutorial

This tutorial showcases a real-time AI agent built with the NVIDIA NeMo Agent Toolkit, powered by NVIDIA Nemotron models, controlling a Reachy Mini Robot. The agent uses an intelligent LLM router to dynamically route between:

• Nemotron nano text for text-based interactions
• Nemotron nano VLM (Vision Language Model) for visual understanding
• REACT agent for tool-based actions

Architecture

The system consists of three main components running in parallel:

1. Reachy Mini Daemon - Controls the robot hardware (or simulation)
2. Bot Service - Processes vision and speech, coordinates robot actions
3. NeMo Agent Service - Handles AI agent logic with intelligent routing between models

Prerequisites

• Python 3.10+
• uv package manager
• NVIDIA API Key (for Nemotron models)
• ElevenLabs API Key (for text-to-speech)

Setup Instructions

1. Clone and Navigate to Repository

cd /path/to/ces-tutorial

2. Create Environment File

Create a .env file in the main directory with your API keys:

NVIDIA_API_KEY=your_nvidia_api_key_here
ELEVENLABS_API_KEY=your_elevenlabs_api_key_here

3. Setup Bot Service

In a terminal window:

cd bot
uv venv
uv sync

4. Setup NeMo Agent Service

In a separate terminal window:

cd nat
uv venv
uv sync

Running the System

You'll need three terminal windows running simultaneously.

Terminal 1: Start Reachy Mini Daemon

Navigate to the bot directory and start the robot daemon:

For macOS:

cd bot
uv run mjpython -m reachy_mini.daemon.app.main --sim --no-localhost-only

For Linux:

cd bot
uv run -m reachy_mini.daemon.app.main --sim --no-localhost-only

Note: The --sim flag runs the robot in simulation mode. Remove it if using actual hardware.

Terminal 2: Start Bot Service

In the bot directory:

cd bot
uv run --env-file ../.env python main.py

This service handles:

• Vision processing through the robot's camera
• Speech recognition and text-to-speech
• Robot movement coordination
• Emotional expression through dance moves

Terminal 3: Start NeMo Agent Service

In the nat directory:

cd nat
uv run --env-file ../.env nat serve --config_file src/ces_tutorial/config.yml --port 8001

This launches the NeMo Agent Toolkit server with intelligent model routing capabilities.

How It Works

1. Vision & Audio Input: The bot captures visual information and listens for speech
2. Agent Processing: The NeMo Agent router intelligently selects the appropriate model:
   • Text queries → Nemotron nano text model
   • Visual queries → Nemotron nano VLM
   • Action requests → REACT agent with tool calling
3. Robot Actions: Based on the agent's response, the bot executes movements, expressions, or speaks

Demo

Check out ces_tutorial.mp4 to see the system in action!

Project Structure

ces-tutorial/
├── bot/                    # Robot control and vision/speech processing
│   ├── main.py            # Main bot orchestration
│   ├── nat_vision_llm.py  # Vision and LLM integration
│   └── services/          # Robot services (moves, speech, etc.)
├── nat/                    # NeMo Agent Toolkit configuration
│   └── src/ces_tutorial/
│       ├── config.yml     # Agent configuration
│       └── functions/     # Router and agent implementations
└── .env                   # API keys (create this file)

Troubleshooting

• Port conflicts: Ensure port 8001 is available for the NeMo Agent service
• API key errors: Verify your .env file is properly formatted and contains valid keys
• Robot connection issues: Check that the Reachy daemon started successfully before launching the bot service

Resources

• NVIDIA NeMo Agent Toolkit
• Reachy Mini Robot
• NVIDIA Nemotron Models