Core Concepts
Factors influencing evolutionary dynamics in joint evolution of robot body and control.
Abstract
An investigation delves into the complexities of jointly optimizing robot design and controller evolution. The study focuses on physically creating robots with a rich morphological space, including voxel-based chassis, wheels, legs, and sensors. It explores the challenges of matching controllers to designs and evolving closed-loop control. The research emphasizes the importance of understanding the intertwined 'evolution+learning' processes for high-performing robots. Factors like synchronous vs asynchronous evolution, replacement mechanisms, and reward mechanisms are studied to determine their impact on performance. Results indicate that asynchronicity combined with goal-based selection yields the highest performance.
The study builds upon previous work in evolutionary robotics by considering both body and brain co-optimization. It highlights the difficulty of evolving designs with closed-loop controllers that utilize sensory information effectively. Researchers aim to evolve diverse robots that are high-performing while exploring a wide range of morphologies to avoid convergence to suboptimal solutions. Various frameworks exist for integrating learning algorithms with evolutionary processes using different architectures.
The proposed algorithm involves nested optimization processes where an evolutionary algorithm optimizes robot morphology while a learning algorithm optimizes controller behavior for each morphology produced. Modifications to existing methods enhance evolutionary operators' flexibility and parallelization efficiency inspired by prior works in the field.
Stats
Asynchronous update combined with goal-based selection leads to highest performance.
Removal of worst performing members enhances overall performance.
Novelty-based objective function does not consistently improve exploration compared to goal-based objectives.
Quotes
"Results show that asynchronicity combined with goal-based selection and a ‘replace worst’ strategy results in the highest performance."
"The combination of (asynchronous, replace-worst, goal-based) evolution leads to the highest performing robots."