DeepMimic: learning simulated movement from example clips

What changed

A simulated character can reach a target while moving in an implausible way. DeepMimic combines a task reward with a reward for matching a reference motion, so the policy learns both what to accomplish and how to move. Random starting points within a motion and early termination of unsuccessful attempts help training reach difficult parts of a skill. The work demonstrates dynamic movements and several ways to combine skills. Its contribution is this practical imitation-and-task learning approach, using existing reinforcement-learning machinery rather than introducing PPO itself.

Scientific scope

A simulated character can reach a target while moving in an implausible way. DeepMimic combines a task reward with a reward for matching a reference motion, so the policy learns both what to accomplish and how to move. Random starting points within a motion and early termination of unsuccessful attempts help training reach difficult parts of a skill. The work demonstrates dynamic movements and several ways to combine skills. Its contribution is this practical imitation-and-task learning approach, using existing reinforcement-learning machinery rather than introducing PPO itself. The demonstrations are manually prepared short motion clips, with retargeting and control settings. The reported characters are simulated; an Atlas-shaped model is not deployment on a real Atlas robot. Training can take days per skill, and larger motion libraries or arbitrary transitions are not demonstrated. Phase-based motion limits retiming, finite perturbation tests do not guarantee robustness, and normalized rewards are not necessarily success probabilities or physically attainable maxima.

Evidence

The linked original works support the scoped statement and contribution attribution. WRS has not independently replayed experiments or reconstructed a full mathematical proof. The demonstrations are manually prepared short motion clips, with retargeting and control settings. The reported characters are simulated; an Atlas-shaped model is not deployment on a real Atlas robot. Training can take days per skill, and larger motion libraries or arbitrary transitions are not demonstrated. Phase-based motion limits retiming, finite perturbation tests do not guarantee robustness, and normalized rewards are not necessarily success probabilities or physically attainable maxima.

Contributors

• The named authors in the research record · SCIENCE: Source-listed human researchers and engineers for the exact method or resource. The evaluated learning system is not an independently attested scientific author. Earlier methods, funding and data-collection roles remain separate. Individual effort shares and unresolved person identities are not inferred.

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

Provisional WRS assessment, edition 2026-10-23.6. Independent expert calibration is pending. This grade does not certify correctness or assign ownership.

Full scientific record · Browse contribution grades · Request a correction