A3I Project - Choreography Planning by Decoupling A* Deterministic Search and Heuristic Computation

Overview

This project uses advanced techniques to design and evaluate dance choreographies for the NAO robot. By combining heuristic methods, a custom A* search algorithm, and large language model (LLM) evaluations, the system generates creative and engaging robot movements tailored to specific constraints and criteria.

Features

• Choreography Evaluation: Rates choreography based on storytelling, rhythm, movement technique, public involvement, space use, human characterization, and human reproducibility.
• A Search Algorithm*: Custom implementation to find the optimal sequence of dance moves while adhering to constraints like duration and mandatory moves.
• Heuristic Computation: Combines past scores (g-function) and future estimations (h-function) to guide the search process.
• Dynamic Scoring: Allows adjusting weights for different scoring components (e.g., waypoint adherence, duration bonus).

Components

1. LLM Scoring Functions:
   • g_llm: Evaluates the coolness of a complete dance sequence.
   • h_llm: Estimates the coolness of potential moves for an incomplete sequence.
2. Custom A Implementation*:
   • Uses LLM-generated scores to find optimal paths for robot choreography.
   • Incorporates dynamic constraint satisfaction.
3. Scoring Criteria:
   • gh_weight: Importance of coolness scores.
   • t_weight: Importance of adhering to waypoints.
   • db_weight: Bonus for optimal duration.
   • csc_weight: Importance of satisfying constraints.

Input Data

• Initial State: Starting position of the robot.
• Final State: Target position to complete the choreography.
• Waypoints: Intermediate positions the robot must pass through.
• Mandatory Positions: Specific moves required in the choreography.
• Intermediate Positions: Optional moves for enhanced creativity.
• Graph: A connectivity graph representing valid transitions between moves.

Outputs

• A complete sequence of moves.
• Coolness scores for the choreography.
• Computation cost and number of LLM calls.

Installation

1. Clone this repository:

   git clone https://github.com/lorenzobalzani/llm-a-star-planner.git

2. Make sure to have conda installed onto your system.
3. Set up the conda environment by launching the script ./setup_env.sh in the terminal.
4. Activate it with conda activate a3i.
5. For installing the ipython kernel, please run python -m ipykernel install --user --name=a3i.

Setup

Create a file named secrets.env in the root directory of the project and add the following lines:

AZURE_OPENAI_ENDPOINT = <your-deployment-edpoint>
AZURE_OPENAI_API_KEY = <your-api-key>
OPENAI_API_VERSION = <your-api-version>

If you prefer to use an LLM other than Azure OpenAI, update the LLM client definitions in the notebook accordingly

Example Output

Choreography: ['INITIAL_stand_init', 'mandatory_stand', 'diagonal_left', ... 'FINAL_crouch']
Cost: 0.06451€
Number of calls to LLM: 59
Elapsed time: 31.50s

Contributing

Contributions are welcome! Please open an issue or submit a pull request.

Authors

• Davide Bombardi ([email protected])
• Lorenzo Balzani ([email protected])

License

This project is licensed under the MIT License.