Grasping Living Objects With Adversarial Behaviors Using Inverse Reinforcement Learning

Abstract

Living objects are difficult to grasp since they can actively elude capture by adopting adversarial behaviors that are extremely hard to model or predict. In this case, an inappropriately strong contact force may hurt the struggling living objects and a grasping algorithm that can minimize the contact force whenever possible is required. To solve this challenging task, in this article, we present a reinforcement-learning (RL)-based algorithm with two stages: the pregrasp stage and the in-hand stage. In the pregrasp stage, the robot focuses on the living object's adversarial behavior and approaches it in a reliable manner. In particular, we use inverse RL to encode the living object's adversarial behavior into a reward function. The negative value of the learned reward function is then used to train a high-quality grasping policy that can compete with the living object's adversarial behavior with the RL framework. In the in-hand stage, we use RL to train a grasp policy such that the dexterous hand can grab the living object with the minimal force. A set of dense rewards are also specifically designed to encourage the robot to grasp and hold the living object persistently. To further improve the grasp performance, we explicitly take into account the structure of the dexterous robot hand by treating the hand as a graph and adopting graph convolutional network to formulate the grasping policy. We conduct a set of experiments to demonstrate the performance of our proposed method, in which the robot can grasp living objects with the success rate of 90% and 95% in the pregrasp and in-hand stages, respectively. The contact force applied by the robotic hand to the living object is dramatically reduced in comparison with the baseline grasping policy.