Lifelong Active Inference of Gait Control

This script implements the gait control learning method proposed in the paper Lifelong Active Inference of Gait Control, which learns and adapts its gait control in real-time while interacting with the environment. The method employs two cognitive models for motor control: Active Inference Control and Internal Model (IM) ensemble.

The robot processes sensory measurement with its World Model to infer action achieving a given sensory reference. Over time, the method autonomously decides when to expand its World Model by learning a new IM which allows the robot to adapt its gait control to new situations.

The method is demonstrated in a simulated experiment with a hexapod robot that has paralyzed legs. The robot decides to extend its World Model by learning the Paralysis model resulting in paralysis compensation.

After a time, the robot recovers from paralysis and perceives sensory input similar to that before the paralysis thus the robot selects the Walking model and recalls walking with all legs.

The proposed method was deployed on a real hexapod robot, where the robot was able to learn and adapt its gait control in real time.

Quickstart

Requirements

• Tested on Ubuntu 20.4.
• Python>=3.8
• venv

Installation

1. Install the software and open the virtual environment:

git clone git@github.com:comrob/laicg.git
cd laicg
python3 -m venv venv # create virtual environment
source venv/bin/activate # activate virtual environment
pip install -r requirements.txt # install requirements into virtual environment

Note: The venv (virtual environment) is installed in the laicg folder.

2. The robot is simulated in CoppeliaSim which should be installed on your machine. Tested versions: 4.3.0 (rev. 10.), 4.6.0 (rev. 18.).

wget "https://downloads.coppeliarobotics.com/V4_6_0_rev18/CoppeliaSim_Player_V4_6_0_rev18_Ubuntu20_04.tar.xz"
tar -xf CoppeliaSim_Player_V4_6_0_rev18_Ubuntu20_04.tar.xz
cd CoppeliaSim_Player_V4_6_0_rev18_Ubuntu20_04
./coppeliaSim -f ../resources/coppeliasim/_plane.ttt

Execution

3. Open new terminal in the project root: Run the experiment and its evaluation with the following commands:

source venv/bin/activate # activate virtual environment
python -m paralysis_recovery_test.run my_new_experiment test --transfer example
python -m paralysis_recovery_test.run my_new_experiment eval

The full experiment and evaluation typically takes around 30 minutes to complete. After the experiment, the collected data are postprocessed, which can require up to 8GB RAM. Performance may vary based on machine specs

Results

4. The eval command generates files in paralysis_recovery_test/results/reports.
   1. 
   2. 

On the right, log-odds evolution ( zero-model:best model (black), paralysis:walking (cyan)) where the value lesser than zero (below the red dash line) indicates that the former is less likely than the latter. On the left, the resulting path of the robot is going towards the goal location [-100,-100].

iii.

Comparison of Walking and Paralysis internal models. The normalized parameters show the sensorimotor coupling strength. The Paralysis IM has low coupling strength for the middle left (L2) and rear right (R3) legs.

iv. Expected results computed from recorded data:

Acknowledgments

This work was supported by the Czech Science Foundation (GAČR) under research project No. 21-33041J.

Cite

If you use this code in your research, please cite the following paper:

@article{laicg2025,
  author={Szadkowski, Rudolf and Faigl, Jan},
  journal={IEEE Transactions on Neural Networks and Learning Systems}, 
  title={Lifelong Active Inference of Gait Control}, 
  year={2025},
  volume={},
  number={},
  pages={1-12},
  doi={10.1109/TNNLS.2025.3579814}
 }