DEEP REINFORCEMENT LEARNING FOR DANCE POSE OPTIMIZATION

Abstract [English]

Deep Reinforcement Learning (DRL) has become an effective framework of sequential decision-making in high-dimensional and complex control problems, but little has been done to apply it to expressive human movement. This work is a unified DRA algorithm to optimize a dance pose, which an intelligent agent is trained to produce the smooth and stable dance pose and aesthetically compose a pose in a simulated kinematic environment. Generation of dance poses is modeled formally as a Markov Decision Process with incorporation of joint level kinematics, time related dependencies and balance constraints in the state space and pose corrections expressed as a continuous control action. The given framework incorporates pose estimation results and biomechanical constraints to guarantee physical feasibility and motion synthesis that is safe of injuries. Several types of DNR algorithms are tested such as Deep Q-Networks, Proximal Policy Optimization, and Actor-Critic variants to determine their appropriateness to fine-grained pose refinement. The pose accuracy, motion smoothness, energy efficiency and balance stability are well coordinated in a reward function that allows a multi-objective optimization that is in line with technical correctness and artistic quality. Curriculum learning is used to bring a gradual complexity to the poses so that the agent is able to move on to dynamic dance patterns. Substantial experimental investigation shows that policy-gradient-related techniques are more convergent stable and realistic than value-based baselines.