MAP-Elites Enables Powerful Stepping Stones and Diversity for Modular Robotics

Area of Science:

• Robotics
• Artificial Intelligence
• Evolutionary Computation

Background:

• Modular robots offer adaptability through reconfiguration, but optimizing morphology and control simultaneously is challenging.
• Interdependencies between morphology and control can lead optimization algorithms to local optima, hindering progress.
• Existing optimization methods struggle with the complex search space of modular robot design.

Purpose of the Study:

• To compare the effectiveness of three Evolutionary Algorithms in optimizing modular robot morphologies and controllers.
• To evaluate algorithms based on performance, morphological diversity, and adaptability to new environments.
• To investigate the role of diversity in achieving high-performing solutions in modular robotics.

Main Methods:

• Comparison of two objective-based search algorithms (with and without diversity promotion) against the Quality Diversity algorithm MAP-Elites.
• Evaluation of algorithm performance on evolving robot morphologies and controllers for locomotion tasks.
• Analysis of genealogical ancestry and population transfer to assess diversity and adaptability.

Main Results:

• MAP-Elites evolved the highest performing solutions and generated the greatest morphological diversity.
• MAP-Elites demonstrated superior performance recovery when populations were transferred to more difficult environments.
• Analysis revealed MAP-Elites creates more diverse and effective 'stepping stones' for optimization.

Conclusions:

• MAP-Elites is highly suitable for morphology-control co-optimization in modular robotics.
• Morphological diversity is crucial for adaptability and performance in changing environments.
• Diverse ancestral pathways generated by MAP-Elites correlate with maximum task performance.