easagru

Code and data for the paper The value of information and the emergence of communication in evolved e-puck robots (arXiv:2610.13118).

Two e-puck robots, rebuilt in PyBullet from the official Webots model and validated against it, must find a food patch they can only detect with their ground sensors. They can emit a one-bit signal whose bearing and strength the partner perceives. Each robot is controlled by a Res-GRU (GRU with a residual input-output path), alone or as one channel of the EASA basal ganglia architecture. The fitness only measures the time spent on the food.

pip install -r requirements.txt

Python 3.9+ tested. Results in the paper were obtained on one machine (Mac mini M1): physics is not bit-identical across platforms, so a re-run on another machine gives statistically equivalent, not identical, trajectories.

python3 easagru_stats.py

Writes results_tables/: Table 3 (small arena), Table 4 (evolved basal ganglia priorities), Table 5 (big arena) and statistical_tests.txt (paired Wilcoxon tests, Welch t-test between conditions).

One folder per evolution run (condition_seed; _big = evolved in the 1.5 m arena): run_config.json (all settings, library versions), evolution_log.csv (one row per generation), champion.npy (the evaluated genome), evaluation_<run>.csv (100 test episodes x signal normal/off/random), diagnosis_<run>.txt (behaviour and signal specificity). runs/robustness/ holds the robustness test. Checkpoints and per-generation genomes are not included (size); every run can be regenerated with the commands below.

python3 easagru_standalone.py --brain gru --genome runs/gru_s1/champion.npy
python3 easagru_standalone.py --brain gru --task big --genome runs/gru_s1/champion.npy
python3 easagru_standalone.py --brain easa_evo --genome runs/easa_evo_s1/champion.npy
python3 easagru_standalone.py --brain gru_compass --genome runs/gru_compass_s1/champion.npy

Use the run's condition as --brain and --task big for runs evolved in the big arena. Add --signal off or --signal random to watch the signal ablations. Keys: SPACE fast mode, P pause, T sensor rays, N channel labels, L console log, F camera follows robot A, R new episode, Q quit. The front LED turns yellow while a robot emits. Every step is logged to easagru_log.csv (disable with --nolog).

python3 test_step1_kinematics.py

Checks the e-puck against the Webots kinematics: straight-line speed (expected 0.1256 m/s), spin rate on its axis (expected 4.83 rad/s) and tilt at rest. All experiments use physics at 120 Hz, with the controller acting every 8 steps (66.7 ms). On the machine used for the paper this gave 0.1255 m/s and 4.62 rad/s, against 0.1253 m/s and 4.65 rad/s at 240 Hz.

To repeat the check at 240 Hz, set in easagru_config.json

"timestep": 0.00416667,
"control_period_steps": 15

run the test, and restore the original values (0.00833333 and 8) afterwards. The remaining validation tests (sensors, environment, basal ganglia) run with scripts/run_tests.sh.

scripts/run_tests.sh                          # validation of robot, sensors, environment, BG (~6 min)
scripts/run_evolution.sh gru 1                # one evolution (~6 h on a Mac mini M1, 6 workers)
scripts/run_evolution.sh gru 1 big            # big arena (~9 h)
python3 easagru_evaluate.py --run gru_s1 --use-champion --episodes 100
scripts/diagnose_run.sh gru_s1
scripts/robustness.sh
scripts/signal_value_big.sh

Conditions: gru, gru_compass, easa_safe, easa, easa_evo. With a freshly evolved run (which has best_genomes/), scripts/evaluate_run.sh <run> --episodes 100 also selects the champion as in the paper.

• easagru_world.py, easagru_sensors.py, easagru_signal.py, easagru_env.py: robot, sensors, one-bit signal, two-robot foraging task (PyBullet).
• easagru_bg.py: EASA basal ganglia core (checked against easa_ref/bg_core.py).
• easagru_brain.py: Res-GRU, innate channels, EASA variants, conditions.
• easagru_evolve.py: genetic algorithm (homogeneous team, implicit fitness).
• easagru_evaluate.py, easagru_diagnose.py, easagru_robustness.py, easagru_signal_value.py, easagru_stats.py: evaluation and tables.
• easagru_standalone.py, easagru_log.py: GUI viewer and its step log.
• robots/: e-puck URDF (generated by make_epuck_urdf.py); webots_ref/: the Webots protos all measurements come from (Apache 2.0).

If you use this work, please cite the paper:

Montes-Gonzalez, F. (2026). The value of information and the emergence of communication in evolved e-puck robots. arXiv:2610.13118. https://doi.org/10.48550/arXiv.2610.13118

@misc{montesgonzalez2026value,
  author        = {Montes-Gonzalez, Fernando},
  title         = {The value of information and the emergence of communication in evolved e-puck robots},
  year          = {2026},
  eprint        = {2610.13118},
  archivePrefix = {arXiv},
  primaryClass  = {cs.RO},
  doi           = {10.48550/arXiv.2610.13118}
}

If you use the code or data specifically, please also cite the archived version:

Montes-Gonzalez, F. (2026). easagru: emergence of a one-bit signal in two e-puck robots with a Res-GRU and EASA basal ganglia (Version 1.0.0) [Software]. Zenodo. https://doi.org/10.5281/zenodo.23051976

This project is released under the MIT License.

The e-puck model files in the webots_ref folder are unmodified copies from the Webots project by Cyberbotics Ltd., distributed under the Apache License 2.0.