Deep Imitation Learning of Sequential Fabric Smoothing From an Algorithmic Supervisor

摘要: Sequential pulling policies to flatten and smooth fabrics have applications from surgery manufacturing home tasks such as bed making folding clothes. Due the complexity of fabric states dynamics, we apply deep imitation learning learn that, given color (RGB), depth (D), or combined color-depth (RGBD) images a rectangular sample, estimate pick points pull vectors spread maximize coverage. To generate data, develop simulator an algorithmic supervisor that has access complete state information. We train in simulation using domain randomization dataset aggregation (DAgger) on three tiers difficulty initial randomized configuration. present results comparing five baseline learned report systematic comparisons RGB vs D RGBD inputs. In simulation, achieve comparable superior performance analytic baselines. 180 physical experiments with da Vinci Research Kit (dVRK) surgical robot, trained attain coverage 83% 95% depending tier, suggesting effective smoothing can be sensing is valuable addition alone. Supplementary material available at https://sites.google.com/view/fabric-smoothing.