“DeepMimic: Example-Guided Deep Reinforcement Learning of Physics-Based Character Skills,” by Xue Bin Peng, Pieter Abbeel, Sergey Levine, and Michiel van de Panne, was published at SIGGRAPH 2018 (and posted to arXiv in April 2018). It tackled a long-standing problem in character animation: motion-capture clips look great but are rigid, while physics-simulated characters react naturally to pushes and uneven ground but are hard to make move convincingly. DeepMimic combines both.
The method trains a control policy with deep reinforcement learning to imitate a reference motion clip while running inside a physics simulation. The character is rewarded for matching the pose of the example motion frame by frame, which keeps the movement looking natural, while the physics engine forces it to actually balance and respond to forces. The paper shows characters performing “a broad range of example motion clips,” including “highly-dynamic actions such as motion-captured flips and spins” - backflips, spinkicks, cartwheels - and recovering after being shoved.
DeepMimic became a widely cited reference for physics-based character animation and active ragdoll control. The same author later extended the idea (in follow-ups such as AMP and ASE) toward characters that learn large libraries of reusable skills, and the approach influenced both game animation and humanoid robotics, where imitating a reference motion under real physics is exactly the control problem to solve.