SwarmChat is an innovative project that enables intuitive communication with swarm robotics through natural language. This system integrates advanced audio transcription, text processing, and safety mechanisms with a live simulation environment that visualizes a swarm of agents executing behavior trees.

🚀 This project is Funded by the European Union’s UTTER programme, in collaboration with the UTTER consortium.

• Audio Input Processing:

  • Record commands via a microphone.
  • Translate speech into English using the facebook/seamless-m4t-v2-large model.
  • Perform a safety check on the translated text before execution.

• Text Input Processing:

  • Enter text commands for swarm control.
  • Translate text using EuroLLM (EuroLLM-9B-Instruct).
  • Detect unsafe or inappropriate content with an integrated safety module.

• Safety Module:

  • Utilizes Llama-Guard model (Llama-Guard-3-8B) for safety classification.
  • Identifies unsafe content across predefined categories (e.g., violent crimes, privacy violations, hate speech).
  • Ensures commands comply with safety standards.

• Swarm Simulation:

  • Visualize a swarm of agents in a live simulation powered by Violet simulator and Pygame.
  • Agents are controlled by behavior trees defined in an XML file (tree.xml), using the py_trees framework.
  • Real-time simulation updates streamed via a Gradio web interface.

• Behavior Tree Generator:

  • Falcon3-10B-Instruct-BehaviorTree model to dynamically generate behavior trees in XML format.
  • Automatically extracts available behaviors from the SwarmAgent class and constructs a detailed prompt using a predefined XML template.
  • Generates and saves new behavior tree configurations (updating tree.xml) based on user-specified tasks.

• Integrated Interface:

  • A unified Gradio web interface for both audio and text inputs.
  • Live streaming of the simulation environment.
  • Seamless switching between different input modalities.

• Backend:

  • Python
  • Transformers (Hugging Face)
  • PyTorch
  • Pygame
  • Threading and Queue modules for simulation management

• Frontend:

  • Gradio for an interactive web-based interface.

• AI Models:

  • Speech Processing: Seamless-m4t-v2-large for audio transcription and translation.
  • Text Processing: EuroLLM-9B-Instruct for text translation.
  • Safety Classification: Llama-Guard-3-8B for content safety assessment.
  • Behavior Tree Generation: Falcon3-10B-Instruct-BehaviorTree for creating and updating behavior trees.

• Behavior Trees:

  • Agents utilize behavior trees—parsed from XML and built with py_trees—to dictate their actions within the simulation.

1. Clone the repository:

   git clone https://github.com/Inventors-Hub/SwarmChat.git
   cd SwarmChat

2. Install dependencies:

   pip install -r requirements.txt

3. Setup AI Models:

• Place the EuroLLM model file (EuroLLM-9B-Instruct-Q4_K_M.gguf) at the specified path in text_processing.py.
• Place the LLaMA Guard model file (llama-guard-3-8b-q4_k_m.gguf) at the specified path in safety_module.py.
• Place the DeepSeek model file (Falcon3-10B-Instruct-BehaviorTree-3-epochs-GGUF) at the specified path in bt_generator.py.

4. Run the Application:

   python app.py

5. Access the Interface:

   Open your browser and navigate to https://127.0.0.1:7860 to start using SwarmChat.

• app.py
  The main application integrates audio/text processing, behavior tree generation, and the live simulation. It sets up the Gradio interface, handles simulation streaming, and routes user inputs to the appropriate processing modules.

• speech_processing.py
  Implements audio transcription and translation using the facebook/seamless-m4t-v2-large model.

• text_processing.py
  Translates text commands using EuroLLM (EuroLLM-9B-Instruct).

• safety_module.py
  Utilizes LLaMA Guard to assess the safety of incoming commands, ensuring compliance with safety policies.

• bt_generator.py
  Dynamically generates behavior trees in XML format by extracting behaviors from the SwarmAgent class, constructing a prompt, and querying Falcon3-10B-Instruct-BehaviorTree model. The generated XML is saved to tree.xml for simulation use.

• simulator_env.py
  Powers the simulation environment, manages agent behaviors using XML-defined behavior trees, and handles real-time simulation updates.

This work was funded by the European Union under the UTTER programme.
We gratefully acknowledge the support of the entire UTTER consortium.