SwarmAdaptivePattern_GECCO25

This project implements the evolutionary algorithm CMA-ES to evolve the optimal robot controller (D), enabling each robot of a swarm to reproduce a phenotypic pattern at the group scale, matching the visual display target (A). The swarm of moving robots is modeled in an original sliding-puzzle setup (C) — a grid where each cell can either be occupied by an autonomous robot or remain empty. This distributed system is characterized by two parameters: density, which refers to the number of robots relative to the total available positions in the grid; and fluidity, which represents the probability of an agent moving to a nearby empty position at each time step. Communication between robots is local, occurring through signal exchanges in a neighbor-to-neighbor von Neumann topology (D).

To cite this work (BibTeX):
@inproceedings{Loi2025,
author = {Alessia Loi and Nicolas Bredeche},
title = {Evolving Neural Controllers for Adaptive Visual Pattern Formation by a Swarm of Robots},
booktitle = {Proceedings of GECCO 2025 @ Málaga (hybrid), The Genetic and Evolutionary Computation Conference (GECCO '25)},
year = {2025},
publisher = {ACM},
address = {New York, NY, USA},
pages = {1--9}
}

Acknowledgement: This work is supported by the Agence Nationale pour la Recherche under Grant ANR-24-CE33-7791

To run an experiment: Learning, Swarm (generalization), and Plots

• Set the learning and generalization parameters by modifying learning_params.json and swarm_params.json. The parameters are already settled to reproduce the paper results.
• In the Bash script launch.sh, uncomment the lines that you wish to run. If you run learning_main.py only, you can later execute swarm_main.py or the analysis scripts by commenting out learning_main.py and filling the learning_analysis_dir and swarm_analysis_dir fields.
• From the src directory, execute the Bash script launch.sh: ~/SwarmAdaptivePattern_GECCO25/src$ ./launch.sh
• The results will be stored in a newly created simulationAnalysis folder.

Details

• In the function plot_flag in environments.py: the plot of the signal pattern does not work for grid_nb_rows ≤ 10 and grid_nb_cols ≤ 10 because signal values can be negative. Easy fix: comment out the special debug display for small grids.
• To reproduce the paper results, set:
  • "env_name": "sliding_puzzle_incremental"
  • "sliding_puzzle_incremental_switch_eval": 0
  • "sliding_puzzle_incremental_nb_deletions_units": null
  • "sliding_puzzle_incremental_nb_deletions_percent": [0.0, X], where X is the complement of density (i.e., 1.0 - density).
• In learning_params.json, the parameter sliding_puzzle_incremental_nb_deletions_percent corresponds to the complement of density (1.0 - density). Example 1: If density = 1.0 (meaning all grid positions are occupied by robots), then sliding_puzzle_incremental_nb_deletions_percent = 0.0. Example 2: If density = 0.8 (meaning 20% of the grid positions are empty), then sliding_puzzle_incremental_nb_deletions_percent = 0.2.