Evolution of adaptive synapses: robots with fast adaptive behavior in new environments

Area of Science:

• Robotics
• Artificial Intelligence
• Computational Neuroscience

Background:

• Evolving neural controllers often results in fixed parameters, limiting adaptability.
• Robots require robust adaptation mechanisms for real-world, dynamic environments.

Purpose of the Study:

• To evolve mechanisms for parameter self-organization in neural controllers, rather than evolving parameters directly.
• To assess the adaptation capabilities of these self-organizing controllers in novel environments.

Main Methods:

• Encoding local adaptation rules for synaptic plasticity within neural controllers.
• Allowing robots equipped with these controllers to move freely in diverse environments.
• Measuring performance in environments significantly different from those used during evolution.

Main Results:

• Evolutionary adaptive controllers significantly outperformed standard fixed-weight controllers in speed and task success.
• The proposed method demonstrated scalability to large neural network architectures.
• Controllers successfully adapted to new sensory characteristics, including sim-to-real transfer and cross-platform deployment.

Conclusions:

• Evolving self-organizing mechanisms offers a superior approach to fixed parameters for neural controllers.
• This method provides robust adaptation capabilities crucial for real-world robotic applications.
• The approach is effective across various environmental changes and robotic platforms.