Automatic Acquisition of Actions for Animated Agents

The generation of animated human figures especially in crowd scenes has many applications in such domains as the special effects industry, computer games or for the simulation of the evacuation from crowded areas. Automation in action creation eliminates the need for human labour, which shortens the task of generation of the crowd scenes and also reduces the costs. This paper addresses a shortcoming of the architectures designed for creation of animated scenes with autonomous agents by proposing a module for automatic acquisition of new high-level actions. Agents use reinforcement learning to acquire these actions and the chosen algorithm is the deterministic version of Q-learning. This allows for easy definition of the task, since only the ultimate goal of the learning agent must be defined. Generated actions can then be used to enrich the animation produced by the animation system. The paper also compares results achieved when training agents with forward and inverse kinematics control.