System and method for using imitation learning in training robot control strategies

Implementations described herein relate to training and refining robot control strategies using imitation learning techniques. The robot control strategy may be initially trained based on human presentations of various robot tasks. In addition, the robot control strategy may be refined based on human intervention when the robot performs the robot task. In some implementations, the robot control strategy may determine whether the robot will fail in the execution of the robot task and prompt a human to intervene in the execution of the robot task. In additional or alternative implementations, a representation of a sequence of actions may be visually appeared to be presented to a person, and the person may actively intervene in execution of a robotic task. 本文描述的实现方式涉及使用模仿学习技术来训练和细化机器人控制策略。机器人控制策略可以基于各种机器人任务的人演示来初始训练。此外，机器人控制策略可以基于在机器人执行机器人任务时的人为干预来细化。在一些实现方式中，机器人控制策略可以确定机器人是否将在机器人任务的执行中故障，并且提示人为干预机器人任务的执行。在附加的或替代的实现方式中，动作序列的表示可以在视觉上显现以呈现给人，并且人可以主动干预机器人任务的执行。