TL;DRDTMR jointly optimizes the timing and the control of a reference motion for every control step with sampling-based MPC on a GPU, so legged robots can track dynamic source motions. Evaluated on two hours of human motion with four humanoids and transferred to a real robot.
Legged robots can learn expressive whole-body skills from the motions of humans and animals. Due to the morphology gap between the source and the robot, however, the motion must be tailored to the dynamic properties of the robot. In particular, dynamic motions such as a jump require careful adjustment, since their timing and control are interdependent. We propose dense temporal motion retargeting (DTMR), which jointly optimizes timing and control within a single optimization, where dense means that the timing is adjusted for every control step. This dense formulation enables DTMR to deform only the parts of the motion that need a change in timing. The problem is solved with sampling-based model predictive control (MPC) in parallel on a GPU. We evaluate DTMR against baselines on two hours of human motion with four humanoid robots, where the results show that DTMR outperforms baseline methods, particularly on dynamic motions. We also show that allowing more temporal deformation yields more precise retargeting. We further compare DTMR with a baseline that optimizes the temporal dimension, where the result shows that DTMR retargets more precisely under the same deformation budget while being about 19× faster. Lastly, policies trained on our references transfer to a real humanoid robot.
01 · Real world
Policies trained on DTMR references, on a Unitree G1 and a Booster K1. Real time.
02 · Morphology
Robots differ in size, mass and actuation, so the same motion needs a different timing on each. DTMR optimizes timing and control jointly, per robot.
Swipe to see all five panels →
High jump, one wall clock, real time. Left: the human (SOMA mocap body of a SEED clip, Motion Data by Bones Studio) and the DTMR references of the Unitree G1 and the Booster K1, each on its own timing. Right: the two robots executing policies trained on those references. The taller G1 and the small K1 leave the ground and land at different times from the human and from each other.
03 · Challenge
Retiming a motion changes the control it requires, and the available control limits the timing the robot can achieve.
04 · Approach
The progress through the source motion is one more state and its rate one more action: one sampling-based MPC optimizes timing and control together.
Track the reference subject to the dynamics.
Progress φ becomes a state, its rate an action.
Control and timing in one MPC problem.
The human and the kinematic reference run on the source clock; DTMR runs on its own robot clock, coloured by its local time-scale: blue where the robot moves slower than the source motion, orange where it moves faster. The plot below the robots is the time warp. Switch to Real robot and the DTMR card becomes the real Unitree G1 executing the same motion in real time.
05 · Feasibility
One tracking policy per retargeter, rolled out on its own references. Orange ghost: reference; solid: policy; DTMR in green.
06 · Dense vs. segment-wise
The temporal baseline warps a whole segment with one time-scale; DTMR stretches only where the dynamics require it.
07 · Deformation
One weight sets how much DTMR may deform the timeline: lower means more deformation and better tracking.
08 · Dataset
Two hours of human motion from SEED (Motion Data by Bones Studio) retargeted to four humanoids. Below, a 72-clip sample per robot as physics rollouts in real time; the 3D explorer has all 999 clips.
Every one of the 999 clips, the SEED body next to the four DTMR references. Type to filter; the list is grouped by motion class.
Citation
@misc{kim2026dtmr,
title = {Dense Temporal Motion Retargeting for Legged Robots},
author = {Kim, Jaeryeong and Yoon, Taerim and Cheng, Jin and Choi, Sungjoon and Coros, Stelian},
year = {2026},
eprint = {2609.38617},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2609.38617}
}Acknowledgements
The human motions in this work come from the Skeletal Everyday Embodiment Dataset (SEED), used under the Bones Studio Dataset License Agreement.
Motion Data by Bones Studio (https://bones.studio/)