With adversarial reinforcement learning, physically simulated characters can be developed that automatically synthesize lifelike and responsive behaviors.
A character is first trained to perform complex motor skills by imitating human motion data. Once the character has acquired a rich repertoire of skills, it can reuse those skills to perform new tasks in a natural, lifelike way.
This model then allows you to generate motions for new scenarios, without tedious manual animation or new motion data from real actors.