Lifelong Active Inference of Gait Control.

Area of Science:

• Robotics
• Artificial Intelligence
• Computational Neuroscience

Background:

• Robots face challenges in dynamic environments due to unknown factors.
• Continuous adaptation requires updating robot-environment interaction knowledge.
• Existing methods like predictive coding (PC) and active inference control (AIC) have limitations, including catastrophic forgetting.

Purpose of the Study:

• To propose an autonomously expanding self-verifying world model (WM) for robot adaptation.
• To combine PC with incremental knowledge representation (internal model principle) to overcome catastrophic forgetting.
• To enable model-based gait control in robots for enhanced longevity and performance.

Main Methods:

• Developed an autonomously expanding self-verifying world model (WM).
• Integrated predictive coding (PC) with the internal model (IM) principle for incremental knowledge representation.
• Validated the method in virtual and real scenarios using a hexapod walking robot.

Main Results:

• The proposed method enabled real-time adaptation to leg paralysis and recovery in a hexapod robot.
• The robot demonstrated improved performance and generated novel behaviors.
• The approach outperformed existing state-of-the-art methods in adaptation tasks.

Conclusions:

• The developed world model effectively supports robot adaptation in dynamic and unknown environments.
• The method offers interpretable decisions and knowledge, promising functional scalability.
• This approach enhances robot longevity and performance through continuous self-verification and learning.