Shaping the Evolutionary Dynamics of Robot Morphology via Adaptive Control Learning

Robot co-design typically couples within-lifetime reinforcement learning for controller evaluation with cross-generational morphological evolution via evolutionary algorithms. While previous research established that well-adapted bodies enable faster control learning—a property known as morphological intelligence—the reciprocal effect of how control learning shapes morphological evolution remains largely unexplored. This paper investigates how the duration of control learning influences natural selection and evolutionary dynamics across generations.

The authors analyze morphological contributions to control learning by decoupling them into two orthogonal dimensions: morphological intelligence, defined as the convergence speed of learning, and true potential, representing the ultimate performance ceiling. Using simulated voxel-based soft robots across manipulation and locomotion tasks, the authors evaluate how premature fitness evaluations distort the evolutionary process and propose AdaControl, an adaptive scheduling method that dynamically allocates minimally sufficient control learning to ensure unbiased selection.

The experiments reveal that stopping control learning prematurely systematically underestimates a morphology's true potential. This truncation biases natural selection toward fast learners at the expense of long-term performance, severely restricting design space exploration, optimization efficiency, and morphological diversity. Consequently, the widely reported morphological Baldwin effect emerges as an artifact of this evaluation bias rather than a fundamental biological tendency.

To overcome these limitations, the proposed AdaControl method monitors disproportionate selection for morphological intelligence during evolution and extends learning budgets only when necessary. This simple genetic algorithm combined with AdaControl matches the efficiency of exhaustive control while reducing computation by up to 80%, rivaling expensive generative-model-based co-design methods while discovering a more diverse repertoire of high-performing designs.