Pretrained Dynamics Learning of Numerous Heterogeneous Robots and Gen2Real Transfer

Abstract

Acquiring dynamics is vital for robotic learning and serves as the foundation for planning and control. This article addresses two essential inquiries: How can one develop a model that encompasses a vast array of diverse robotic dynamics? Is it possible to establish a model that alleviates the burdens of data collection and domain expertise necessary for constructing specific robot models? We explore the dynamics present in a dataset containing numerous serial articulated robots and introduce a novel concept, “Gen2Real,” to transfer simulated, generalized models to physical, and specialized robots. By randomizing dynamics parameters, topological configurations, and model dimensions, we generate an extensive dataset that corresponds to varying properties, connections, and quantities of robotic links. A structure adapted from the generative pretrained transformer is employed to approximate the dynamics of a multitude of heterogeneous robots. Within Gen2Real, we transfer the pretrained model to a target robot using distillation to enable real-time computation. The results corroborate the superiority of the proposed method in terms of accurately learning an immense scope of robotic dynamics, managing commonly encountered disturbances, and exhibiting versatility in transferring to distinct robots.