Evolving Robot Morphology with Manipulation Learning Inspired by Human Hand Evolution

The evolution of human hands is intricately linked to their morphologies, behaviors, and selection pressures. This biological evolution process provides critical insights for robot design, facilitating robot evolution through environmental interaction and skill learning. In this context, we propose a Baldwin-effect-inspired framework that endows robots with the capacity for long-term enhancement in response to environmental changes. This framework integrates large-scale population evolution and morphology-aware manipulation learning. Each robotic hand undergoes extensive testing of its grasp and manipulation abilities with various objects. Leveraging this framework, these hand agents validate several prior hypotheses related to morphological embodied intelligence: the relative advantage of human-like segment models in manipulative potential, the impact of the environment on the final evolved morphologies, and the influence of morphology difference on the agent's learning ability. These phenomena are not independent, as multiple agents with different segment models exhibit convergent evolution and morphological specialization. Furthermore, this natural evolutionary capability has been implemented in three-, four-, and five-fingered reconfigurable robotic hands, each following unique evolutionary trajectories that evolve towards higher bilateral symmetry. The evolved prototypes have demonstrated superior performance, surpassing anthropomorphic designs by human experts.