Vision Guided Cutting Skill Acquisition for Cognitive Robots

In this paper, we present a cognitive learning framework for the humanoid robot to acquire cutting skill. Our framework consists of two main modules: 1) A visual module that is trained using the deep feature to visually estimate object firmness level, and 2) A motor skill learning module that adopts reinforcement learning method to generate optimized control parameters for cutting. We propose a method that is based on the consideration of energy efficiency for the cutting motion. Experiments conducted on both tasks show that 1) the visual module is able to estimate firmness level of a novel object with state-of-the-art performance, and 2) the motor learning module, given the estimated firmness level from the visual module, is able to improve the actual performance of cutting on a humanoid robot (Baxter). Also, we show that the energy efficiency measure we proposed is a proper criterion for evaluating the cutting skill on cognitive robots.