← ArchivePaper2018
DeepMimic: Example-Guided Deep Reinforcement Learning of Physics-Based Character Skills
Xue Bin Peng, Pieter Abbeel, Sergey Levine, Michiel van de Panne
Physics-based character controller using deep RL to imitate reference motion clips, achieving diverse athletic skills with natural-looking dynamics.
Abstract
A longstanding goal in character animation is to combine data-driven specification of behavior with a system that can execute a similar behavior in a physical simulation, enabling realistic responses to perturbations and environmental variation. We show that well-known reinforcement learning methods can be adapted to learn robust control policies capable of imitating a broad range of example motion clips, while also learning complex recoveries, adapting to changes in morphology, and accomplishing user-specified goals. The method handles keyframed motions, highly dynamic actions such as motion-captured flips and spins, and retargeted motions. By combining a motion-imitation objective with a task objective, characters can be trained to react intelligently in interactive settings.
How to read this
- Category
- Method: physics-based character control via deep reinforcement learning
- Contributions
- Adapts deep RL to learn control policies that imitate a broad range of reference motion clips in physics simulation
- Combines a motion-imitation objective with a task objective so characters pursue user goals while staying natural
- Handles keyframed, highly dynamic (flips, spins), and retargeted motions, and learns complex recoveries plus adaptation to morphology changes
- Context
- Builds on the authors' earlier deep-RL locomotion work (Peng et al., Terrain-Adaptive Locomotion Skills Using Deep Reinforcement Learning), advancing example-guided imitation of reference motion.Builds on: Terrain-Adaptive Locomotion Skills Using Deep Reinforcement Learning
- Correctness
- Demonstrated across diverse athletic skills with natural-looking dynamics and perturbation recovery; results depend on quality reference clips and per-skill training, and the imitation-plus-task objective balance is central to the behavior obtained.
- Clarity
- Accessible in motivation; a first pass conveys the imitation-plus-task idea, a second pass is needed for the reward design and training setup.
- How to read it
- Focus on the reward formulation (imitation plus task) and how reference clips are used; a second pass pays off for the RL training details if you plan to reproduce or extend it.
Built upon by
- Physics-based Motion Capture Imitation with Deep Reinforcement Learning 2018
- SFV: Reinforcement Learning of Physical Skills from Video 2018
- DReCon: Data-Driven Responsive Control of Physics-Based Characters 2019
- Learning Body Shape Variation in Physics-based Characters 2019
- Scalable Muscle-Actuated Human Simulation and Control 2019
- Model Predictive Control with a Visuomotor System for Physics-based Character Animation 2020
- UniCon: Universal Neural Controller for Physics-Based Character Motion 2020
- AMP: Adversarial Motion Priors for Stylized Physics-Based Character Control 2021
- PFPN: Continuous Control of Physically Simulated Characters using Particle Filtering Policy Network 2021
- Machine Learning Summit: Walk Lizzie, Walk! Emergent Physics-Based Animation through Reinforcement Learning 2022
- MotionBricks: Scalable Real-Time Motions with Modular Latent Generative Model and Smart Primitives 2026
Related work
- SFV: Reinforcement Learning of Physical Skills from Video 2018 / SIGGRAPH Asia
- AMP: Adversarial Motion Priors for Stylized Physics-Based Character Control 2021 / SIGGRAPH
- C·ASE: Learning Conditional Adversarial Skill Embeddings for Physics-based Characters 2023 / SIGGRAPH Asia
- SuperTrack: Motion Tracking for Physically Simulated Characters Using Supervisory Signals 2021 / SIGGRAPH Asia
Keywords
This page summarises the entry and links to its original source. The archive never hosts or redistributes the publication itself.Show it in the full archive list →