Generalization of Real-Time Motion Control with DRL Using Conditional Rewards and Symmetry Constraints

Deep Reinforcement Learning has been increasingly explored as a method for generating physics-based motions in articulated characters. However, effective control tools are still necessary to better guide the learning process and provide animators with greater control and reliability over the resulting animations. This paper proposes new control tools, including the generalization of real-time control, conditional rewards, symmetry constraints, and a user interface. Real-time control allows dynamic adjustment of chosen parameters, conditional rewards simplify the competition between rewards, symmetry constraints reduce uncoordinated movements, and the user interface facilitates training and animation parameter specification. The proposed control tools show promise in improving the quality and control of physics-based character animation.