Predictive Display from Computer Vision Models

Abstract

In tele-manipulation, delays as small as a few tenths of a second can affect performance. Robot operators dissociate their control actions with what they see on delayed video and have to adopt slow move-and-wait strategies or may completely fail to perform high precision tasks. Predictive Display (PD) mitigates this problem by rendering visual feedback that reflects the operator’s motion immediately. Conventional PD is based on a-priori CAD models and calibrations. Using modern computer vision, we have implemented on-line model capture and tracking which allows the rendering of textured graphical PD. We validate different types of PD and compare them to using delayed video in an alignment task. Graphical PD was found comparable to a no-delay situation, while task completion on average took 48% longer with a relatively short 300 ms delay.