Layered Evolution in Robotics: A New Approach for Complex Tasks

Title: Layered Evolution in Robotics: A New Approach for Complex Tasks

Research Question: Can layered evolution, a combination of incremental and modularized evolution with elements from subsumption architecture, improve the performance of robot controllers in complex environments?

Methodology: The study used a simulated robot with the task of learning which light source to approach in an environment with obstacles. The robot's controller was divided into layers, each controlled by an evolutionary neural network. Layers were evolved sequentially, with the lowest layer evolving first and higher layers added as the lowest layer's fitness improved. The fitness function was changed for each layer, and all lower layers were kept fixed during the evolution of each layer.

Results: The results showed that layered evolution performed at least as well as monolithic evolution, incremental evolution, and modularized evolution. The evolved layers were merged back into a single network, and the performance was found to be comparable or better than the individual layers.

Implications: The study suggests that layered evolution may be a superior approach for many tasks in robotics. It allows for a more efficient use of resources and a better handling of complexity, making it a promising method for scaling up evolutionary robotics. The approach may also provide insights for other fields that rely on layered systems and complex tasks.

Link to Article: https://arxiv.org/abs/0405027v1 Authors: arXiv ID: 0405027v1