Improve Data-Efficiency for Deep Imitation Learning

Deep neural networks are powerful tools to learn representations from high-dimensional data. Deep imitation learning has been studied to enable robots to learn unknown dynamics and acquire complex visual-motor skills, which maps the raw pixels of the robot's observations to actions based on a pre-trained model obtained via leveraging the demonstration data from humans. However, learning from raw visual inputs typically requires a large database for model training. A robot cannot perform a new task with reasonable behaviors when given limited demonstration data on the task. That is to say, it is impractical for the robot to acquire skills and adapt to new scenarios for task execution. This is known as one of the key limitations for deep imitation learning, which should be addressed to improve its generality for adaptation to new tasks.