BeyondMimic

Common

Trains a humanoid to reproduce human motion, then distills that into a guided diffusion model for new tasks.

BeyondMimic is a whole-body humanoid control framework released in August 2025 by Koushil Sreenath's group at UC Berkeley and C. Karen Liu's group at Stanford, built in two stages. The first stage is motion tracking: a simple, unified reinforcement-learning recipe learns in simulation to reproduce motions from the LAFAN1 motion-capture dataset, then deploys zero-shot to a real Unitree G1, performing highly dynamic moves like aerial cartwheels, spinning kicks, and sprinting. The second stage distills a chosen set of tracking policies into a single latent-space diffusion model; at inference, sampling is guided with a simple cost function, letting it handle new tasks — waypoint navigation, joystick teleoperation, obstacle avoidance — with no retraining. It pushes “imitating motion” forward into “using motion as a prior for general-purpose control,” and the motion-tracking code is open-sourced.

ExampleThe same diffusion model can switch from “follow the joystick” to “walk around obstacles to a target point” just by swapping in a different cost function.

Also called: BeyondMimic: From Motion Tracking to Versatile Humanoid Control via Guided Diffusion
Related: Motion Tracking · Diffusion Model · DeepMimic · ASAP · Unitree G1 · LAFAN1
Sources: BeyondMimic (arXiv 2508.08241)
BeyondMimic 项目主页 (Chinese)
As of: 2025-11