BeyondMimic: From Motion Tracking to Versatile Humanoid Control via Guided Diffusion

Qiayuan Liao, Takara E. Truong, Xiaoyu Huang, Yuman Gao, Guy Tevet, Koushil Sreenath, C. Karen Liu

Seminal August 2025

Teaser, in which the robot walks in the lab while a person pushes it with a pad.

TL;DR One RL motion-tracking recipe with shared hyperparameters lets a Unitree G1 reproduce cartwheels, spin-kicks, sprints and dances from human motion. A latent diffusion model distilled from these trackers is then steered at test time toward joystick, waypoint, obstacle-avoidance and keyframe tasks it never trained on. In a 77-person study, people judged its walking and running more human-like than Unitree's own controller 70.8% of the time.

What they do

Learning humanoid skills from human motion usually needs per-motion reward tuning, and the resulting trackers can only replay their clips. BeyondMimic trains motion trackers with one compact RL recipe, then distills them into a single latent diffusion model that task costs steer at test time toward unseen tasks. Both stages transfer zero-shot to a Unitree G1.

How it works

Tracking. One PPO policy per reference motion, all with the same reward and hyperparameters. Body poses are tracked relative to an anchor body so that global drift is tolerated. The reward is one uniform tracking term plus three penalties. Domain randomization covers only friction, joint offsets, torso center of mass and random pushes.
Latent space. A VAE trained with DAgger on the trackers learns a 32-D latent, which its decoder turns into joint setpoints.
Guided diffusion. A transformer denoises state-latent trajectories of 4 past and 16 future steps. Gradients of simple costs (velocity error, waypoint distance, an SDF obstacle barrier) steer every denoising step, and future keyframes are inpainted. Inference takes about 20 ms on an onboard RTX 4060 Mobile.

Results

Of the 2.5 h of motion trained in simulation, 30 clips (15 min) ran on hardware. They include outdoor aerial cartwheels, with pelvis angular velocity up to 20 rad/s, and five back-to-back Ronaldo jump-turns.
In a 77-person study, people picked these motions as more human-like than Unitree’s controller 70.8% of the time (84.7% for running).
The guided model tracked velocity commands in simulation with 12.14% error walking and 13.65% running. On hardware it switched from walking to running from a velocity command alone, chained keyframed cartwheels with walking and detoured around an obstacle.
Without the latent space, cartwheel success in MuJoCo drops from 95% to 5%. A 10 ms deployment delay makes tracking fail in two of three trials.

Limitations

The 0.64 s horizon is too short for long-horizon planning. Conditioning on history can trap the model in repetitive motion. Countering that takes larger guidance weights, which make the robot stumble at the start and end of motions. Guidance handles coarse objectives but not fine-grained ones, and its weights still need light tuning. The waypoint and obstacle tasks used motion capture for localization.

Videos

Guided diffusion on hardware, with the inset showing the velocity command (green arrow) and predicted body trajectories.

Motion tracking of Cristiano Ronaldo's jump-turn celebration, repeated back to back without motion capture.

Figures

From Figure 1: (A) Skills learned from human motion by motion tracking, including an aerial cartwheel and kicks in a ... — From Figure 1: (A) Skills learned from human motion by motion tracking, including an aerial cartwheel and kicks in a forest and running on a sports field. (B) One guided diffusion controller handling joystick teleoperation, waypoint navigation, obstacle avoidance and motion inpainting.

From Figure 7: (A) Human motions are tracked with RL using one recipe and shared hyperparameters, with the target re-... — From Figure 7: (A) Human motions are tracked with RL using one recipe and shared hyperparameters, with the target re-anchored to the robot's current pose so it can drift and recover. (B) A VAE trained with DAgger compresses the tracking policies into a latent space, and a transformer diffusion model is trained on state-latent rollouts; at deployment the current latent is decoded into an action. (C) Task-specific objectives guide the model at test time.

From Figure 6: (A) Motion inpainting, where keyframes 0.2 s apart turn walking into a cartwheel, and a long run alter... — From Figure 6: (A) Motion inpainting, where keyframes 0.2 s apart turn walking into a cartwheel, and a long run alternates velocity-commanded walking and running with keyframed cartwheels. (B) Combined waypoint and obstacle-avoidance costs make the robot detour around an obstacle on its way to the goal.

Abstract

The human-like form of humanoid robots positions them uniquely to achieve the agility and versatility in motor skills that humans possess. Learning from human demonstrations offers a scalable approach to acquiring these capabilities. However, prior works either produce unnatural motions or rely on motion-specific tuning to achieve satisfactory naturalness. Furthermore, these methods are often motion- or goal-specific, lacking the versatility to compose diverse skills, especially when solving unseen tasks. We present BeyondMimic, a framework that scales to diverse motions and carries the versatility to compose them seamlessly in tackling unseen downstream tasks. At heart, a compact motion-tracking formulation enables mastering a wide range of radically agile behaviors, including aerial cartwheels, spin-kicks, flip-kicks, and sprinting, with a single setup and shared hyperparameters, all while achieving state-of-the-art human-like performance. Moving beyond the mere imitation of existing motions, we propose a unified latent diffusion model that empowers versatile goal specification, seamless task switching, and dynamic composition of these agile behaviors. Leveraging classifier guidance, a diffusion-specific technique for test-time optimization toward novel objectives, our model extends its capability to solve downstream tasks never encountered during training, including motion inpainting, joystick teleoperation, and obstacle avoidance, and transfers these skills zero-shot to real hardware. This work opens new frontiers for humanoid robots by pushing the limits of scalable human-like motor skill acquisition from human motion and advancing seamless motion synthesis that achieves generalization and versatility beyond training setups.

Summaries are written automatically from each paper's text. Figures and videos come from the paper and its project page; all credit goes to the authors.

Matt Strong