There has been a recent paradigm shift in robotics to data-driven learningfor planning and control. Due to large number of experiences required fortraining, most of these approaches use a self-supervised paradigm: usingsensors to measure success/failure. However, in most cases, these sensorsprovide weak supervision at best. In this work, we propose an adversariallearning framework that pits an adversary against the robot learning the task.In an effort to defeat the adversary, the original robot learns to perform thetask with more robustness leading to overall improved performance. We show thatthis adversarial framework forces the the robot to learn a better graspingmodel in order to overcome the adversary. By grasping 82% of presented novelobjects compared to 68% without an adversary, we demonstrate the utility ofcreating adversaries. We also demonstrate via experiments that having robots inadversarial setting might be a better learning strategy as compared to havingcollaborative multiple robots.
展开▼
