Skip to main content
Back

DeepMimic's Creator Ships MimicKit to Fix Motion AI's Reproducibility Crisis

Xue Bin Peng released a clean reinforcement learning codebase bundling DeepMimic, AMP, ASE and a new diffusion-based motion prior called SMP.

  • MimicKit bundles DeepMimic, AMP, ASE, AWR, LCP, ADD and SMP in one Apache-2.0 codebase.
  • Supports Isaac Gym, Isaac Lab and Newton (MuJoCo Warp) via a single engine config flag.
  • New SMP method uses a frozen diffusion model as a reusable, task-agnostic style reward.
  • SMP aggregates SDS losses over fixed diffusion timesteps to stabilize PPO training.
  • Includes AMASS/SMPL and GMR conversion scripts plus multi-GPU distributed training out of the box.
  • Detailed Starter Guide paper covers the algorithms and implementation choices.

Motion imitation research has a reproducibility problem. Every paper reimplements DeepMimic or AMP slightly differently, small numerical details silently ruin training, and the reference codebases tend to bit-rot around specific simulator versions. MimicKit, from Xue Bin Peng (the author behind DeepMimic and AMP), ships a single lightweight framework with correct implementations of the whole family of methods, plus a new one built on diffusion models.

The repo describes itself as a suite of motion imitation methods for training motion controllers, designed to be clean and lightweight, with minimal dependencies. It is Apache-2.0 licensed, Python-only, and sitting at around 2.1k stars. The associated Starter Guide paper walks through the algorithms in more depth.

Simulated Unitree Go2 trotting alongside a reference motion clip

Every mocap-to-controller method under one roof

Rather than scattering across seven repos, MimicKit exposes a consistent CLI and config system for the classic and current motion imitation algorithms:

  • DeepMimic: the original tracking-reward approach that made humanoids do spinkicks and backflips.
  • AMP (Adversarial Motion Priors) and ASE (Adversarial Skill Embeddings): GAN-style discriminators trained against mocap so a policy learns stylistic naturalness without exact frame tracking.
  • AWR (Advantage-Weighted Regression),