Goal-aware generative adversarial imitation learning from imperfect demonstration for robotic cloth manipulation

Generative Adversarial Imitation Learning (GAIL) can learn policies without explicitly defining the reward function from demonstrations. GAIL has potential to with high-dimensional observations as input, e.g., images. By applying a real robot, perhaps robot be obtained for daily activities like washing, folding clothes, cooking, and cleaning. However, human demonstration data are often imperfect due mistakes, which degrade performance of resulting policies. We address this issue by focusing on following features: (1) many robotic tasks goal-reaching tasks, (2) labeling such goal states in is relatively easy. With these mind, paper proposes Goal-Aware (GA-GAIL), trains policy introducing second discriminator distinguish state parallel first that indicates data. This extends standard framework more robustly desirable even demonstrations through goal-state promotes achieving state. Furthermore, GA-GAIL employs Entropy-maximizing Deep P-Network (EDPN) generator, considers both smoothness causal entropy update, achieve stable learning two discriminators. Our proposed method was successfully applied real-robotic cloth-manipulation tasks: turning handkerchief over clothes. confirmed it learns task-specific design. Video experiments available at URL.