A Learning Based Hierarchical Control Framework for Human–Robot Collaboration

In this paper, using the ball and beam system as an illustration, a control scheme is developed on human-robot collaboration, i.e., a two-level hierarchical framework is proposed to establish a robust human-robot collaboration (HRC) policy. On the high level, a deep reinforcement learning (DRL) algorithm is presented to plan the desired beam rotational velocity. The low level is constructed by a human-intention perception module and a robust collaboration policy design module. For the first module, a probabilistic model is fitted by using the Gaussian process regression (GPR) approach to predict human-hand velocities, and prediction results follow Gaussian distributions where mean values and variances represent predicted human-hand velocities and corresponding prediction confidences, respectively. For the second module, a robust collaboration policy is established by fusing a proactive policy and a conservative policy, where the proactive policy is used to control the robot to achieve the desired beam rotational velocity by using the predicted human-hand velocities. The conservative policy is designed to ensure the collaboration safety. The weighted parameters for fusion are adaptively tuned based on the prediction precision and confidence. Experiments are conducted on controlling ball position on a beam jointly by a human and a robot with vision data, and experimental results show the effectiveness of the designed robust collaboration policy. Note to Practitioners—Predicting human future behaviors and moderating robot behaviors accordingly is a long-standing problem for human-robot collaboration (HRC) tasks, such as assembling, transporting, etc. Existing approaches generally regard human behaviors as noises or only build simple human models without prediction confidence. This paper proposes a learning-based hierarchical framework that will derive a robust and safe HRC policy considering human behaviors, prediction confidence, and task-related optimality. The framework is validated by a representative experiment where human and robot are asked to jointly control a ball and beam system.