Comparing Robot Controller Optimization Methods on Evolvable Morphologies.

Area of Science:

• Robotics
• Artificial Intelligence
• Evolutionary Computation

Background:

• Modular robots require effective gait-learning algorithms for controller optimization.
• Joint evolution of robot morphology and controllers necessitates adaptable learning strategies.
• Evaluating gait-learners without prior morphological knowledge is crucial for generalizability.

Purpose of the Study:

• To compare the efficiency, efficacy, and morphological sensitivity of Bayesian Optimization, Differential Evolution, and Evolution Strategy as gait-learning algorithms.
• To assess the performance of these algorithms on a diverse set of twenty robot morphologies.
• To determine which gait-learning algorithm performs best when applied to novel and varied robot body plans.

Main Methods:

• Utilized Bayesian Optimization, Differential Evolution, and Evolution Strategy as gait-learning algorithms.
• Tested algorithms on a suite of twenty distinct modular robot morphologies.
• Evaluated algorithms based on solution quality (walking speed), number of evaluations, and sensitivity to morphological variations.

Main Results:

• Bayesian Optimization and Differential Evolution achieved comparable solution quality with fewer evaluations than Evolution Strategy.
• Evolution Strategy demonstrated higher sensitivity to morphological differences, leading to more variable performance.
• Evolution Strategy exhibited greater variance in outcomes across repeated runs, indicating higher susceptibility to chance.

Conclusions:

• Bayesian Optimization and Differential Evolution are more efficient and robust gait-learning algorithms for modular robots with unknown morphologies.
• Evolution Strategy's performance is less reliable and more dependent on specific morphological characteristics.
• The choice of gait-learning algorithm significantly impacts performance and adaptability in evolving robotic systems.