CARL: Controllable Agent with Reinforcement Learning for Quadruped Locomotion

摘要

Motion synthesis in a dynamic environment has been a long-standing problemfor character animation. Methods using motion capture data tend to scalepoorly in complex environments because of their larger capturing and labelingrequirement. Physics-based controllers are effective in this regard, albeitless controllable. In this paper, we present CARL, a quadruped agent thatcan be controlled with high-level directives and react naturally to dynamicenvironments. Starting with an agent that can imitate individual animationclips, we use Generative Adversarial Networks to adapt high-level controls,such as speed and heading, to action distributions that correspond to theoriginal animations. Further fine-tuning through the deep reinforcementlearning enables the agent to recover from unseen external perturbationswhile producing smooth transitions. It then becomes straightforward to create autonomous agents in dynamic environments by adding navigationmodules over the entire process. We evaluate our approach by measuringthe agent’s ability to follow user control and provide a visual analysis of thegenerated motion to show its effectiveness.