Learning Expressive Humanoid Locomotion from Monocular Runway Videos

Kyrylo Kolesnichenko1,2 · Irvin Steve Cardenas1 · Jong-Hoon Kim1

1 Advanced Telerobotics Research Laboratory, Kent State University2 Vilnius University

Abstract

Runway walking uses posture, stride, and foot placement to present clothing and express a particular style. However, humanoid robots in fashion shows usually use walking policies designed mainly for balance and speed. We present an end-to-end pipeline that converts a monocular runway video into a deployable humanoid policy through motion recovery, robot retargeting, motion correction, policy training, simulation testing, and physical deployment. We evaluate the pipeline on the Booster K1 using a runway-style motion. The learned policy completed all 20 physical trials without falling and reproduced narrow foot placement together with coordinated leg, torso, and arm movements. Its measured step width ranged from −0.8 to 1.8 cm, compared with 5.1 to 11.4 cm for the standard K1 walk, and included occasional crossover steps.

From video to a running policy

Four panels: a person walking on a runway, the recovered 3D body mesh, the retargeted K1 robot in simulation, and the physical Booster K1 walking.
(A) Monocular runway video  →  (B) recovered 3D human motion  →  (C) retargeted motion in simulation  →  (D) physical Booster K1.

  1. Motion recovery

    GVHMR lifts a single static-camera video into world-grounded SMPL-X body motion.

  2. Ground correction

    Foot contacts snap the motion to a level floor and remove root drift without changing joint angles.

  3. Retargeting

    GMR maps the human skeleton onto the Booster K1’s joints.

  4. Policy training

    A BeyondMimic tracking task is generated and trained in Isaac Lab.

  5. Sim-to-sim

    The exported policy is checked in MuJoCo with the deployment controller.

  6. Hardware

    The same policy runs on the K1 with identical actuator limits, stiffness and damping.

Results on the physical robot

20/20

physical trials completed without a fall, about 23 steps each along a 5 m runway.

Step width between consecutive foot placements

−20 24 68 1012 step width (cm) · negative = crossover Standard K1 walk n = 12 steps 5.1 11.4 Catwalk policy n = 23 steps −0.8 1.8

The learned policy walks with much narrower foot placement than the stock gait, and sometimes crosses over.

Run it yourself

The repository wires pinned GVHMR, GMR and Booster checkouts into one script. Needs x86_64 Linux, Conda and an NVIDIA GPU. 

git clone --recursive https://github.com/Kyk0/expressive-motion.git
cd expressive-motion
bash install.sh                 # creates the conda environments

conda activate expressive-motion
python scripts/process.py inputs/videos/clip.mp4 --isaac-env env_isaaclab
python scripts/make_task.py --clip clip

Full setup, body-model download steps and troubleshooting are in the README.

Citation

@misc{kolesnichenko2026expressive,
  title  = {Learning Expressive Humanoid Locomotion
            from Monocular Runway Videos},
  author = {Kolesnichenko, Kyrylo and Cardenas, Irvin Steve
            and Kim, Jong-Hoon},
  year   = {2026},
  url    = {https://github.com/Kyk0/expressive-motion}
}

Poster

Conference poster for the project

Acknowledgements

This work builds directly on: