Emergence of human-comparable balancing behaviours by deep reinforcement learning

This paper presents a hierarchical framework based on deep reinforcement learning that naturally acquires control policies that are capable of performing balancing behaviours such as ankle push-offs for humanoid robots, without explicit human design of controllers. Only the reward for training the neural network is specifically formulated based on the physical principles and quantities, and hence explainable. The successful emergence of human-comparable behaviours through the deep reinforcement learning demonstrates the feasibility of using an AI-based approach for humanoid motion control in a unified framework. Moreover, the balance strategies learned by reinforcement learning provides a larger range of disturbance rejection than that of the zero moment point based methods, suggesting a research direction of using learning-based controls to explore the optimal performance.