Generalização do controle de movimento em tempo real com drl usando recompensas condicionais e restrições de simetria

Abstract

Producing natural, physics-based movements for articulated characters is a challenging problem. Deep Reinforcement Learning (DRL) has been explored as a solution, but its isolated use still results in unnatural movements and requires significant adjustments. Providing appropriate new control tools is necessary to better guide the learning process and allow the animator to have more control and reliability over the resulting animation. This work addresses this challenge by proposing tools to facilitate the training process and provide new forms of control. The proposed tools consist of the generalization of real-time control, conditional rewards, symmetry constraints, and a user interface adapted to the proposed generalizations. The generalization of real-time control allows any parameter to be adjusted dynamically, increasing flexibility. Conditional rewards introduce tolerances so that conditions can be met simultaneously, simplifying the competition between rewards and avoiding the need for precise weight adjustments. Symmetry constraints reduce the action space, preventing uncoordinated movements, and offer an additional layer of control. A user interface demonstrates these generalizations, allowing training and animation parameters to be specified more easily without the need for programming. The proposed control tools have proven to be useful and promising, according to results obtained that were compatible with the choices made by the interface and that were able to adapt to the environment.