New study uses VLM subjective scores to evolve soft robot animats

In a new paper accepted by Artificial Life and Robotics (arXiv:2608.07537), researchers Shota Miyazaki, Takaya Arita, and Reiji Suzuki introduce an evolutionary framework that replaces traditional fitness functions with subjective evaluations from a Vision-Language Model (VLM). They evolve virtual soft robots—called animats—with flexible morphologies and locomotion capabilities. For each generation, the VLM is shown sequence images of two individuals' movements and asked to compare them using subjective adjectives such as 'adorably' or 'weirdly.' These pairwise comparisons are fed into a genetic algorithm as selection pressure, simultaneously shaping both body plans and gaits without any predefined objective metric.

Experiments show that VLM-based subjective selection accelerates population convergence compared to random selection, and each evaluation term produces distinctive morphological and behavioral traits—e.g., 'adorable' animats evolve compact, endearing forms while 'weird' ones produce unusual, unpredictable motion. In a human auxiliary study, participants' pairwise choices only partly matched the VLM's, yet the overall morphological and locomotion tendencies were qualitatively similar, and repeated evaluations caused noticeable fatigue. Interestingly, similar evolutionary outcomes emerged across different adjectives, suggesting the VLM does not apply terms literally but decomposes them into shared internal evaluation criteria. This work offers a foundational platform for mapping subjective language onto embodied phenotypes and opens new avenues for evolutionary computation, artificial life, and interpretable AI judgment.