Cutaway View Learning for Visually-Guided Assembly with Crane

Object manipulation for construction assembly using a crane is a control problem with highly challenging dynamics, merging contact-rich manipulation, high dynamics uncertainty, and an underactuated system. Learning a vision-guided controller for such a system using reinforcement learning is a promising but challenging approach, as mating surfaces are occluded during the last stage of assembly, making feedback indirect. We present a novel form of cutaway-view privileged information for assembly tasks that is used within a student-teacher framework, making alignment information readily available during the initial stage of training. This is paired with a pretrained encoder and embedding buffer that leverages nonphysical manipulation within the simulator to collect its training data. We evaluate our method on four different assembly-type placement tasks, and find that our system significantly outperforms both kinodynamic planning and standard reinforcement-learning baselines. We also evaluate the ability of our trained controllers to transfer to a realistic simulation environment with different underlying dynamics, demonstrating continued superior performance under deployment with a significant dynamics gap.