When robots learn and inherit: Lamarckian evolution works in predictable change

Co-evolving robot bodies and controllers faces a fundamental coupling problem: morphology constrains what control strategies work, while control determines morphology performance. This work combines evolutionary optimization of body structure with lifetime learning of controllers, using Lamarckian inheritance to pass learned parameters from parent to offspring. Testing on virtual soft robots with two learning methods (Bayesian optimization and reinforcement learning), the authors show Lamarckian inheritance beats pure Darwinian evolution except when environmental changes are both conflicting and unpredictable. Adding sensors to detect environmental changes restores Lamarckian benefits by enabling robots to generalize their learned behaviors across conditions.