Kinodynamic Motion Retargeting for Humanoid Locomotion via Multi-Contact Whole-Body Trajectory Optimization

Xiaoyu Zhang; Steven Haener; Varun Madabushi; Maegan Tucker

cs.ROMar 10, 2026

Kinodynamic Motion Retargeting for Humanoid Locomotion via Multi-Contact Whole-Body Trajectory Optimization

Authors: Xiaoyu Zhang, Steven Haener, Varun Madabushi, Maegan Tucker

Organizations: Georgia Institute of Technology, Atlanta, GA, USA

Abstract

We present the KinoDynamic Motion Retargeting (KDMR) framework, a novel approach for humanoid locomotion that models the retargeting process as a multi-contact, whole-body trajectory optimization problem. Conventional kinematics-based retargeting methods rely solely on spatial motion capture (MoCap) data, inevitably introducing physically inconsistent artifacts, such as foot sliding and ground penetration, that severely degrade the performance of downstream imitation learning policies. To bridge this gap, KDMR extends beyond pure kinematics by explicitly enforcing rigid-body dynamics and contact complementarity constraints. Further, by integrating ground reaction force (GRF) measurements alongside MoCap data, our method automatically detects heel-toe contact events to accurately replicate complex human-like contact patterns. We evaluate KDMR against the state-of-the-art baseline, GMR, across three key dimensions: 1) the dynamic feasibility and smoothness of the retargeted motions, 2) the accuracy of GRF tracking compared to raw source data, and 3) the training efficiency and final performance of downstream control policies trained via the BeyondMimic framework. Experimental results demonstrate that KDMR significantly outperforms purely kinematic methods, yielding dynamically viable reference trajectories that accelerate policy convergence and enhance overall locomotion stability. Our end-to-end pipeline will be open-sourced upon publication.

Figures & tables

Fig. 1: Given a source human motion and measured ground-reaction forces (GRFs), the proposed KDMR performs contact-aware trajectory optimization to generate a contact-consistent and dynamically feasible robot reference. A downstream RL policy tracks this reference on the real humanoid robot.

Fig. 2: Overview of KDMR. Human pose trajectories and GRF measurements are used to infer the contact schedule and initialize the robot motion. The initial trajectory is then refined through contact- and dynamics-constrained whole-body optimization to produce a dynamically feasible robot reference.

Panel of Figure 3 — Fig. 3: Comparison of retargeted trajectories produced by KDMR and GMR over a time segment (20–27 s) of the full motion sequence. (a) Robot base configuration $\bm{\mathchar 29041}_{\mathchar 29026}$ , consisting of three translational and three rotational coordinates. (b) Corresponding base velocity $\dot{\bm{\mathchar 29041}}_{\mathchar 29026}$ . (c) Joint configuration $\bm{\mathchar 29041}_{\mathrm{\mathchar 29036\mathchar 29029\mathchar 29031}}$ of one leg, ordered from the hip-pitch joint to the ankle-roll joint. (d) Corresponding joint velocity $\dot{\bm{\mathchar 29041}}_{\mathrm{\mathchar 29036\mathchar 29029\mathchar 29031}}$ . Here, $\mathchar 29041_{\mathchar 29026\mathchar 24891\mathchar 29033}$ and $\dot{\mathchar 29041}_{\mathchar 29026\mathchar 24891\mathchar 29033}$ denote the $\mathchar 29033$ -th base coordinate and its velocity, while $\mathchar 29041_{\mathrm{\mathchar 29036\mathchar 29029\mathchar 29031}\mathchar 24891\mathchar 29033}$ and $\dot{\mathchar 29041}_{\mathrm{\mathchar 29036\mathchar 29029\mathchar 29031}\mathchar 24891\mathchar 29033}$ denote the position and velocity of the $\mathchar 29033$ -th leg joint, respectively, for $\mathchar 29033\mathchar 12349\mathchar 28720\mathchar 24891\ldots\mathchar 24891\mathchar 28725$ .

		$\mathchar 28997_{\mathrm{\mathchar 29031\text{-}\mathchar 29026\mathchar 29040\mathchar 29029}}$ ( $\mathrm{\mathchar 29037\mathchar 29037}$ )				$\mathchar 28997_{\mathrm{\mathchar 29026\mathchar 29040\mathchar 29029}}$ ( $\mathrm{\mathchar 29037\mathchar 29037}$ )				$\mathchar 28997_{\mathrm{\mathchar 29034\mathchar 29040\mathchar 29029}}$ ( $\mathrm{\mathchar 29037\mathchar 29042\mathchar 29025\mathchar 29028}$ )
Motion	Method	Mean	Median	Min.	Max.	Mean	Median	Min.	Max.	Mean	Median	Min.	Max.
Walk	GMR	14.7	14.6	12.9	17.9	12.8	12.7	10.8	16.0	40.6	40.8	38.5	42.7
Walk	KDMR	9.1	9.1	7.5	12.9	8.3	8.2	6.6	12.3	29.4	29.4	27.7	31.7
Twister	GMR	2 2 . 3 5	2 1 . 8 8	1 9 . 2 5	2 7 . 3 2	1 8 . 2 2	1 7 . 9 9	1 5 . 4 7	2 3 . 4 9	4 6 . 3 1	4 6 . 2 1	4 3 . 3 1	5 1 . 6 2
Twister	KDMR	11.9	11.7	9.8	14.04	10.8	10.5	8.4	14.8	33.6	33.5	31.8	36.3

TABLE I: Comparison of KDMR and GMR across mean errors: global body position error (

\mathchar 28997_{\mathchar 29031\mathchar 8704\mathchar 29026\mathchar 29040\mathchar 29029}

), body position error (

\mathchar 28997_{\mathchar 29026\mathchar 29040\mathchar 29029}

), and joint position error (

\mathchar 28997_{\mathchar 29034\mathchar 29040\mathchar 29029}

). Each error is computed over 100 rollouts with different seeds.

Fig. 4: The height of heel and toe with contact window.

Fig. 5: Mean episode rewards over three independent training runs. Solid lines show the mean, and shaded regions show one standard deviation.

Method	Motion	Runtime [s]
GMR	Walking	7
GMR	Twist	10
KDMR	Walking	88.6
KDMR	Twist	282.4

TABLE II: Runtime comparison of GMR and KDMR for generating retargeted reference trajectories of 800 frames

Explore similar work

ReActor: Reinforcement Learning for Physics-Aware Motion Retargeting
May 7, 2026David Müller, Agon Serifi, Sammy Christen +3Cross-Embodiment Robot Learning Motion Retargeting
Disney Research, Switzerland
Direct Dynamic Retargeting for Humanoid Imitation Learning from Videos
May 22, 2026Constant Roux, Ludovic De Matteïs, Armand Jordana +4Humanoid Robot Control Motion Retargeting
LAAS-CNRS, Université de Toulouse, CNRS, Toulouse, France · IRT Saint-Exupéry, Toulouse, France · Artificial and Natural Intelligence Toulouse Institute (ANITI), Toulouse, France
Shooting for Contact: Contact-Implicit Multiple Shooting for Dynamic Motion Retargeting
Aug 4, 2026Sergio A. Esteban, Jason H. K. Siu, Derrick Mach +4Trajectory Optimization Multibody Dynamics
Department of Mechanical and Civil Engineering, California Institute of Technology, Pasadena, CA, USA. · School of Computing, DePaul University, Chicago, IL, USA.

May 7, 2026J/K move · Enter open · S save

ReActor: Reinforcement Learning for Physics-Aware Motion Retargeting

David Müller, Agon Serifi, Sammy Christen, Ruben Grandia+2

Disney Research, Switzerland

Cross-Embodiment Robot Learning Motion Retargeting

Open paper AskPDF arXiv

Retargeting human kinematic reference motion onto a robot's morphology remains a formidable challenge. Existing methods often produce physical inconsistencies, such as foot sliding, self-collisions, or dynamically infeasible motions, which hinder downstream imitation learning. We propose a bilevel optimization framework that jointly adapts reference motions to a robot's morphology while training a tracking policy using reinforcement learning. To make the optimization tractable, we derive an approximate gradient for the upper-level loss. Our framework requires only a sparse set of semantic rigid-body correspondences and eliminates the need for manual tuning by identifying optimal values for a parameterization expressive enough to preserve characteristic motion across different embodiments. Moreover, by integrating retargeting directly with physics simulation, we produce physically plausible motions that facilitate robust imitation learning. We validate our method in simulation and on hardware, demonstrating challenging motions for morphologies that differ significantly from a human, including retargeting onto a quadruped.

Abstract

Explore similar work

ReActor: Reinforcement Learning for Physics-Aware Motion Retargeting

Direct Dynamic Retargeting for Humanoid Imitation Learning from Videos

Shooting for Contact: Contact-Implicit Multiple Shooting for Dynamic Motion Retargeting