VLM-Driven Genetic Algorithm Evolves Soft Robots with Subjective Selection

A recent paper published on arXiv (2608.07537) introduces a novel framework that incorporates subjective assessments from a Vision-Language Model (VLM) into the genetic algorithm's fitness evaluation and selection stages. The focus of the research is on virtual soft robots characterized by adaptable morphologies and movement capabilities. The VLM analyzes sequences of images depicting the movements of two individuals, with selection occurring through pairwise comparisons using subjective descriptors such as 'adorably' and 'weirdly.' This method applies selection pressure within the genetic algorithm, facilitating the concurrent evolution of both morphology and locomotion. Results indicate that VLM-driven subjective selection enhances population convergence compared to random methods, yielding unique morphologies and motions aligned with each descriptor. An additional experiment involving human participants supported these findings. The paper is classified as cross-type on arXiv, highlighting its interdisciplinary significance, bridging artificial intelligence, evolutionary computation, and subjective aesthetics, with potential implications for robotics and digital art.