GraspAgent 1.0: Adversarial Continual Dexterous Grasp Learning

Grasp is at the core of robotic manipulation tasks. Nonetheless, most 6-DOF methods resort to a one-time setup via intensive analytics and targeting a predetermined domain. On the other hand, learning and adapting in real environments is of great promise to robotics yet challenging. In this context, this letter presents a grasp learning agent embodied in a dual-arm robot with a gripper and suction ends. Theoretically, GraspAgent is equipped with two adversarial components with the capacity for continual improvement. The first component aims to learn the synthesis of high dexterity grasps using a novel <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Grasp Awareness Generative Adversarial Network</i>. Thanks to the designed set of contrastive objectives, GA-GAN yields a high rate of feasible grasps for clutter scenes. The second component is an <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Adversarial Experience Replay</i>. AER promotes certain attributes during the regular rounds of training. For instance, a quality network is forced to adapt the maturity of the sampler by highlighting regions where the sampler performs well. Further, the quality network of each arm is encouraged to learn different grasp primitives, which reflects on better learning capacity. Finally, Experiments on a Yumi robot reveal a final average success rate of 93% after exploiting 17k+ feedback data. Moreover, We show that grasp sampling with GA-GAN surpasses a set of recent baseline methods.