저널 : JICRS(Journal of Institute of Control, Robotics and Systems), Vol.26, No.5, pp.325-334, 2020. 5.
· 논문제목 : "AugGAIL: Augmented Generative Adversarial Imitation Learning for Robotic Manipulation Tasks"
· 저자 : 정은진, 이석준, 김인철
· 요약 : GAIL is an imitation learning model that has shown successful results for control tasks in continuous state-action space through adversarial learning methods that are similar to GANs. This learning model can obtain an expert-like task policy with fewer trial errors than pure reinforcement learning methods. However, it associates with low data efficiency in that it requires many interactions with the environment in order to reach the expert’s performance level. To overcome this limitation, we propose a novel imitation learning model, AugGAIL. AugGAIL has several important extensions compared to GAIL such as replacing TRPO-based policy updates with more efficient PPO-based ones, incorporating task-oriented rewards into the original reward function, applying a novel data sampling strategy to train the discriminator more efficiently, and pre-training the policy network similar to that of behavioral cloning. In this study, we applied the AugGAIL learning model to several robotic object manipulation tasks with continuous state-action space, including pick-up, pick-and-place, and stack. In various experiments using a simulated Jaco arm robot, AugGAIL showed higher efficiency and performance than other imitation learning models.