Video-driven learning of microgravity locomotion for robot astronauts

Abstract Controlling robot astronauts in microgravity is challenging due to complex floating-base dynamics and the scarcity of motion-capture data. This paper proposes a video-driven pipeline to acquire free-floating mobility skills directly from public astronaut footage, serving as a foundational step toward complex biomechanical control in space. Monocular motion reconstruction and kinematic retargeting are employed to generate physically feasible robot references. These priors guide a Goal-Conditioned Reinforcement Learning (GCRL) framework augmented with Adversarial Motion Priors (AMP) to ensure precise tracking and human-like style. Simulations validate the pipeline, demonstrating the feasibility of utilizing monocular video as a supervision source for space robotics.