DeepMimic is a reinforcement learning framework that trains simulated humanoid characters to imitate motion skills from motion capture data.
The framework addresses the problem of learning physics-based character control by using deep reinforcement learning guided by example motions. It trains a simulated character to reproduce mocap-derived skills through an adversarial learning approach, as demonstrated in accompanying research on both example-guided imitation and adversarial motion priors for stylized character control. The system combines C++ simulation environments built on Bullet physics with Python-based training using TensorFlow, connected through a SWIG wrapper.
The codebase is now deprecated, with the README directing users toward MimicKit for faster and more accessible implementations. Developers considering this repository should be aware that it represents research code accompanying academic papers rather than a maintained production tool. The project suits researchers studying physics-based character animation and reinforcement learning, particularly those interested in motion imitation techniques. The setup requires substantial dependencies including Bullet physics engine, Eigen for linear algebra, OpenGL graphics libraries, and MPI for distributed training across multiple processes.
Development on the project appears to have concluded, with no active maintenance indicated and explicit direction toward successor work. The codebase reflects a research implementation designed to accompany specific published papers rather than an evolving tool.
