DeepMimic

Uses reinforcement learning to make a simulated character imitate motion-capture clips, learning physically realistic flips and martial arts.

DeepMimic was published at SIGGRAPH 2018 by Xue Bin Peng and colleagues at UC Berkeley and the University of British Columbia. Simulated characters trained with reinforcement learning alone often move stiffly and unnaturally. DeepMimic adds an imitation reward: at every step in physics simulation, the character scores higher the closer its pose is to a reference motion clip (such as motion-capture data), and this can be combined with an additional task reward. Two of its training tricks later became standard practice: reference state initialization, which starts each episode from a random frame of the reference motion, and early termination, which ends an episode as soon as the character falls. It taught humanoid characters, the Atlas robot, and even a dinosaur to walk, flip, and perform martial-arts moves. Today's motion-tracking approaches for humanoid robots, such as ASAP and BeyondMimic, continue this same line of thinking.

ExampleGiven a motion-capture clip of a human backflip, a simulated humanoid character trained this way learns to perform the backflip in physics simulation and land on its feet.

Also called

DeepMimic: Example-Guided Deep Reinforcement Learning of Physics-Based Character Skills

Related

Motion Tracking · Reinforcement Learning · Reference State Initialization · Early Termination · Adversarial Motion Priors · Motion Capture

Sources

DeepMimic (arXiv 1804.02717)

As of

2018-04