Robot learning by Single Shot Imitation for Manipulation Tasks

In this work, we present a Programming by imitation for a robotic manipulation system, which can be programmed for various tasks from only a single demonstration. The system is primarily based on the three components: i) scene parsing, ii) action classification, and iii) dynamic primitive shape fitting. All the above modules are developed by leveraging state-of-the-art techniques in 2D and 3D visual perception. The primary contribution of this system paper is an imitationbased robotic system that can replicate highly complex tasks by executing elementary task-specific program templates, thus avoiding extensive and exhaustive manual coding. In addition, we contribute by introducing a primitive shape fitting module by which it becomes easier to grasp objects of various shapes and sizes. To evaluate the system performance, the proposed robotic system has been tested on the task of multiple object sorting and reports 91.8% accuracy in human demonstrated action detection, 76.1% accuracy in action execution, and overall accuracy of 80%. We also examine the proposed system's component-wise performance to demonstrate the efficacy and deployability in industrial and household scenarios.