PhysMoDPO: Physically-Plausible Humanoid Motion with Preference Optimization

Yangsong Zhang; Anujith Muraleedharan; Rikhat Akizhanov; Abdul Ahad Butt; Gül Varol; Pascal Fua; Fabio Pizzati; Ivan Laptev

cs.LGMar 13, 2026

PhysMoDPO: Physically-Plausible Humanoid Motion with Preference Optimization

Authors: Yangsong Zhang, Anujith Muraleedharan, Rikhat Akizhanov, Abdul Ahad Butt, Gül Varol, Pascal Fua, Fabio Pizzati, Ivan Laptev

Organizations: Mohamed Bin Zayed University of Artificial Intelligence (MBZUAI) · LIGM, École des Ponts, IP Paris, Univ Gustave Eiffel, CNRS · École polytechnique fédérale de Lausanne (EPFL)

Abstract

Recent progress in text-conditioned human motion generation has been largely driven by diffusion models trained on large-scale human motion data. Building on this progress, recent methods attempt to transfer such models for character animation and real robot control by applying a Whole-Body Controller (WBC) that converts diffusion-generated motions into executable trajectories. While WBC trajectories become compliant with physics, they may expose substantial deviations from original motion. To address this issue, we here propose PhysMoDPO, a Direct Preference Optimization framework. Unlike prior work that relies on hand-crafted physics-aware heuristics such as foot-sliding penalties, we integrate WBC into our training pipeline and optimize diffusion model such that the output of WBC becomes compliant both with physics and original text instructions. To train PhysMoDPO we deploy physics-based and task-specific rewards and use them to assign preference to synthesized trajectories. Our extensive experiments on text-to-motion and spatial control tasks demonstrate consistent improvements of PhysMoDPO in both physical realism and task-related metrics on simulated robots. Moreover, we demonstrate that PhysMoDPO results in significant improvements when applied to zero-shot motion transfer in simulation and for real-world deployment on a G1 humanoid robot.

Figures & tables

Figure 1: PhysMoDPO generates motions that follow textual instructions while respecting physical constraints. Compared to prior methods, our approach produces motions that remain stable and physically realistic when deployed on the Unitree G1 robot.

Category	Method	Approach	Applied at	No test-time opt.	No Extra trainable module	Dynamics-aware reward
Constraints / Projection	Li et al. [ 20 ]	Optimization	Train/Infer	✗	✓	–
	Rempe et al. [ 38 ]	Optimization	Train/Infer	✗	✓	–
	PhysPT [ 57 ]	Projection	Train	✓	✗	–
	Zhang et al. [ 56 ]	Projection	Infer	✗	✓	–
	PhysDiff [ 54 ]	Projection	Sampling	✓	✓	–
Rewards / Finetune	ReinDiffuse [ 7 ]	Finetuning	Finetune	✓	✓	✗

Table 1: Comparison with previous work. We compare alternative ways of incorporating physics constraints into human motion generation. In contrast to other methods, PhysMoDPO make use of dynamics-aware reward and does not require test-time optimization nor additional trainable modules.

Figure 2: Overview of PhysMoDPO. Given a conditioning signal (text and optional joint controls), we sample multiple motions $X$ from a pretrained generator. A fixed physics-based tracking policy then projects each sample into a simulated trajectory $X^{\prime}$ . We compute physics rewards and task rewards on $X^{\prime}$ , construct preference pairs, and finetune the generator with DPO. This generation–finetuning procedure can be iterated.

Method	MM-Dist $\downarrow$	R@1 $\uparrow$	R@2 $\uparrow$	R@3 $\uparrow$	FID $\downarrow$	Jerk $\downarrow$
Real after simulation	16.02	0.6682	0.8326	0.8952	34.07	35.87
MaskedMimic [ 42 ]	19.73	0.4134	0.5568	0.6305	73.79	66.08
MotionStreamer [ 50 ]	$\underline{{17.17}}$	$\underline{{0.5829}}$	0.7510	0.8310	$\underline{{49.14}}$	$\underline{{46.75}}$
SFT	17.23	0.5779	$\underline{{0.7651}}$	$\underline{{0.8355}}$	49.22	48.30
PhysMoDPO	$\mathbf{{16.95}}$	$\mathbf{{0.5853}}$	$\mathbf{{0.7726}}$	$\mathbf{{0.8517}}$	$\mathbf{{48.29}}$	$\mathbf{{43.60}}$

Table 2: Evaluation of text-driven human motion generation with SMPL robot simulation on HumanML3D [ 6 ] dataset. We evaluate MaskedMimic [ 42 ] , MotionStreamer [ 50 ] and PhysMoDPO with text-conditioned generation setting as in MotionStreamer [ 50 ] . The best results are in bold, and the second best results are underlined.

		HumanML3D					OMOMO
	Method	Err. $\downarrow$	MM-Dist $\downarrow$	R@3 $\uparrow$	FID $\downarrow$	Jerk $\downarrow$	Err. $\downarrow$	MM-Dist $\downarrow$	R@3 $\uparrow$	FID $\downarrow$	Jerk $\downarrow$
	Real after simulation	0.0536	3.150	0.7607	0.98	62.96	0.0485	6.679	0.1576	0.59	65.49
	MaskedMimic [ 42 ]	0.2493	5.149	0.4932	3.99	106.49	0.2140	$\mathbf{{6.639}}$	0.1081	5.99	119.52
Original	OmniControl [ 52 ]	0.1998	4.238	0.6123	5.82	115.12	0.3989	7.267	0.1315	20.59	161.95
	SFT	0.1536	3.580	0.6875	2.63	92.56	0.2866	7.218	0.1419	13.28	130.69
	PhysMoDPO	0.1298	3.333	0.7168	0.93	62.31	0.1900	6.947	$\mathbf{{0.1536}}$	3.86	101.88

Table 3: Evaluation of spatial-text human motion controllability with SMPL character control. Left: HumanML3D [ 6 ] . Right: OMOMO [ 15 ] . We evaluate MaskedMimic [ 42 ] , OmniControl [ 52 ] under two training settings and PhysMoDPO with cross-control setting as in OmniControl [ 52 ] . The best results are in bold, and the second best results are underlined.

Figure 3: Visual comparison with SMPL simulation. On top, we compare MaskedMimic [ 42 ] , MotionStreamer [ 50 ] and PhysMoDPO on text-to-motion generation task on HumanML3D [ 6 ] dataset. At the bottom, we show visual results of spatial-text control task on HumanML3D [ 6 ] (left) and OMOMO [ 15 ] (right) dataset. Red balls are the input spatial control signals and red boxes highlight the samples which do not follow the control or lose balance.

Method	M2T $\uparrow$	R@1 $\uparrow$	R@2 $\uparrow$	R@3 $\uparrow$	FID $\downarrow$	Jerk $\downarrow$
Real after simulation	0.8282	0.5731	0.7583	0.8475	0.1201	87.76
MaskedMimic [ 42 ]	0.7156	0.3258	0.4762	0.5761	0.3673	$\mathbf{{83.58}}$
MotionStreamer [ 50 ]	$\underline{{0.7904}}$	$\underline{{0.4673}}$	$\mathbf{{0.6620}}$	$\underline{{0.7558}}$	$\underline{{0.3033}}$	95.08
SFT	0.7869	0.4586	0.6444	0.7430	0.3124	97.99
PhysMoDPO	$\mathbf{{0.7919}}$	$\mathbf{{0.4707}}$	$\underline{{0.6596}}$	$\mathbf{{0.7640}}$	$\mathbf{{0.3029}}$	$\underline{{90.14}}$

Table 4: Evaluation of text-driven human motion generation with G1 robot on HumanML3D [ 6 ] dataset. We evaluate MaskedMimic [ 42 ] , MotionStreamer [ 50 ] and PhysMoDPO with text-conditioned generation setting as in MotionStreamer [ 50 ] . The best results are in bold, and the second best results are underlined.

		HumanML3D					OMOMO
	Method	Err. $\downarrow$	M2T $\uparrow$	R@3 $\uparrow$	FID $\downarrow$	Jerk $\downarrow$	Err. $\downarrow$	M2T $\uparrow$	R@3 $\uparrow$	FID $\downarrow$	Jerk $\downarrow$
	Real after simulation	0.1894	0.7810	0.7241	0.1287	160	0.1730	0.4985	0.1111	0.1099	184
	MaskedMimic [ 42 ]	0.3388	0.6470	0.3958	0.4626	162	0.3157	0.4635	0.0970	0.4324	198
Original	OmniControl [ 52 ]	0.5954	0.7025	0.5504	0.3655	315	0.5951	0.4773	0.1301	0.5007	459
	SFT	0.4606	0.7335	0.6108	0.3047	250	0.4857	0.4836	0.1304	0.4571	369
	PhysMoDPO	0.2918	0.7414	0.6335	0.3090	142	0.3499	0.5011	0.1291	0.3816	239

Table 5: Evaluation of human motion controllability with G1 robot. Left: HumanML3D [ 6 ] . Right: OMOMO [ 15 ] . We evaluate MaskedMimic [ 42 ] , OmniControl [ 52 ] and PhysMoDPO with cross-control setting as in OmniControl [ 52 ] . The best results are in bold, and the second best results are underlined.

Figure 4: Visual results on Unitree G1 robot. Our deployed motion model enables the robot to move in a physically-realistic way while following input text instructions.

Figure 5: User study. Comparison of real-robot motion sequences generated by PhysMoDPO, MaskedMimic [ 42 ] and OmniControl [ 52 ] . PhysMoDPO outperform both competitors in terms of text adherence, motion smoothness and overall stability.

Table 6: Ablation studies. We train OmniControl [ 52 ] and study the impact of the number of iterative training round and different rewards on HumanML3D [ 6 ] dataset. The best results are in bold.

Appendix figures & tables6 assets

Supplementary material from the paper’s appendix.

Appendix

Method	Spatial controllability			Text consistency		FID $\downarrow$	Jerk $\downarrow$
Method	Err. $\downarrow$	Traj err 0.5 $\downarrow$	Traj err 0.2 $\downarrow$	M2T $\uparrow$	R@3 $\uparrow$	FID $\downarrow$	Jerk $\downarrow$
Real after simulation	0.2406	0.2734	0.8105	0.5915	0.3916	0.7976	203.6
MaskedMimic [ 42 ]	0.5173	$\underline{{0.5752}}$	0.9629	0.5459	0.2227	1.132	$\underline{{191.4}}$
OmniControl [ 52 ]	0.5121	0.7129	0.9639	0.5693	0.2803	1.038	279.3
SFT	$\underline{{0.3504}}$	0.6113	$\underline{{0.9443}}$	$\mathbf{{0.5846}}$	$\underline{{0.3574}}$	$\underline{{0.9523}}$	231.4
PhysMoDPO	$\mathbf{{0.2497}}$	$\mathbf{{0.4951}}$	$\mathbf{{0.9141}}$	$\underline{{0.5830}}$	$\mathbf{{0.3643}}$	$\mathbf{{0.8820}}$	$\mathbf{{172.7}}$

Appendix

Table 7: Evaluation of zero-shot human motion controllability with H1 robot on HumanML3D [ 6 ] dataset. We apply the models trained with SMPL simulation and then perform zero-shot evaluation for Unitree H1 robot. The best results are in bold, and the second best results are underlined.

Figure 6: Visual comparison with H1 robot. We transfer the motion generated by MaskedMimic [ 42 ] , OmniControl [ 52 ] and PhysMoDPO to H1 robot and then run Hover [ 11 ] to track the movement (blue points). See also the website.

Control	Data rep	Spatial controllability			Text consistency		FID $\downarrow$	Jerk $\downarrow$
Control	Data rep	Err. $\downarrow$	Traj err 0.5 $\downarrow$	Traj err 0.2 $\downarrow$	MM-Dist $\downarrow$	R@3 $\uparrow$	FID $\downarrow$	Jerk $\downarrow$
Pelvis	HML3D	0.0724	0.1055	0.2920	$\mathbf{{3.0137}}$	0.7900	0.3619	$\mathbf{{33.30}}$
Pelvis	SMPL	$\mathbf{{0.0352}}$	$\mathbf{{0.0420}}$	$\mathbf{{0.1855}}$	3.0255	$\mathbf{{0.7822}}$	$\mathbf{{0.2107}}$	34.38
Left hand	HML3D	0.1320	0.2197	0.6504	$\mathbf{{2.9725}}$	$\mathbf{{0.8037}}$	0.3053	$\mathbf{{30.92}}$
Left hand	SMPL	$\mathbf{{0.0662}}$	$\mathbf{{0.1094}}$	$\mathbf{{0.4756}}$	3.0054	0.7842	$\mathbf{{0.1881}}$	33.37
Right hand	HML3D	0.1317	0.2188	0.6572	$\mathbf{{2.9093}}$	0.7939	0.2523	$\mathbf{{31.60}}$

Appendix

Table 8: Ablation of data representation. We compare OmniControl [ 52 ] model trained on different data representations. Numbers are calculated without applying tracking policy. The best results are in bold.

Table 9: Ablation studies. We train OmniControl [ 52 ] and study the impact of the data scale and preference pair selection strategies on HumanML3D [ 6 ] dataset. The best results are in bold.

$\lambda_{\text{SFT}}$	$\beta$	Spatial controllability			Text consistency		FID $\downarrow$	Jerk $\downarrow$
$\lambda_{\text{SFT}}$	$\beta$	Err. $\downarrow$	Traj err 0.5 $\downarrow$	Traj err 0.2 $\downarrow$	MM-Dist $\downarrow$	R@3 $\uparrow$	FID $\downarrow$	Jerk $\downarrow$
0	20	0.2348	0.5918	0.9023	4.21	0.6133	5.28	116.77
1	20	0.1771	0.5469	0.8887	4.00	0.6387	4.31	103.64
2	20	$\mathbf{{0.1421}}$	$\mathbf{{0.4238}}$	$\mathbf{{0.8418}}$	$\mathbf{{3.36}}$	$\mathbf{{0.7246}}$	$\mathbf{{1.17}}$	$\mathbf{{72.13}}$
5	20	$\underline{{0.1440}}$	$\underline{{0.4463}}$	$\underline{{0.8496}}$	$\underline{{3.40}}$	$\underline{{0.7139}}$	$\underline{{1.82}}$	$\underline{{81.11}}$
10	20	0.1476	0.4629	0.8662	3.49	0.7041	2.29	88.07

Appendix

Table 10: Ablation study on hyperparameters. Here

\lambda_{\text{SFT}}

is the weight of SFT loss on win sample,

\beta

is the temperature parameter from Diffusion-DPO [ 46 ] . The best results are in bold, and the second best results are underlined.

Figure 7: User study design. We show the text input with 2 videos generated by PhysMoDPO and one baseline method. The participant is asked to evaluate the videos according to text-motion similarity, smoothness and stability.

Human Motion Generation

Human motion generation creates articulated movement from language, audio, video, spatial constraints, or interaction goals.

See all 291 papers

RoboForge: Physically Optimized Text-guided Whole-Body Locomotion for Humanoids
Mar 18, 2026Xichen Yuan, Zhe Li, Bofan Lyu +4Humanoid Robot Locomotion Human Motion Generation
MARS Lab, Nanyang Technological University
MoLingo: Motion-Language Alignment for Text-to-Human Motion Generation
Dec 15, 2025Yannan He, Garvita Tiwari, Xiaohan Zhang +4Latent Diffusion Models Text-to-Motion Generation
University of Tübingen, Germany · Tübingen AI Center, Germany · Max Planck Institute for Informatics, Saarland Informatics Campus, Germany +2
TEXEDO : Test Time Scaling for Controller-aware Language-conditioned Humanoid Motion Generation
Jun 22, 2026Jianuo Cao, Yuxin Chen, Yuzhen Song +3Humanoid Robot Control Test-Time Scaling
Nanjing University · University of California, Berkeley · Southern University of Science and Technology

Mar 18, 2026J/K move · Enter open · S save

RoboForge: Physically Optimized Text-guided Whole-Body Locomotion for Humanoids

Xichen Yuan, Zhe Li, Bofan Lyu, Kuangji Zuo+3

MARS Lab, Nanyang Technological University

Humanoid Robot Locomotion Human Motion Generation Language-Conditioned Robot Control Diffusion Models for Motion Generation Synthetic Data Generation for Robotics

Open paper AskPDF arXiv

While generative models have become effective at producing human-like motions from text, transferring these motions to humanoid robots for physical execution remains challenging. Existing pipelines are often limited by retargeting, where kinematic quality is undermined by physical infeasibility, contact-transition errors, and the high cost of real-world dynamical data. We present a unified latent-driven framework that bridges natural language and whole-body humanoid locomotion through a retarget-free, physics-optimized pipeline. Rather than treating generation and control as separate stages, our key insight is to couple them bidirectionally under physical constraints.We introduce a Physical Plausibility Optimization (PP-Opt) module as the coupling interface. In the forward direction, PP-Opt refines a teacher-student distillation policy with a plausibility-centric reward to suppress artifacts such as floating, skating, and penetration. In the backward direction, it converts reward-optimized simulation rollouts into high-quality explicit motion data, which is used to fine-tune the motion generator toward a more physically plausible latent distribution. This bidirectional design forms a self-improving cycle: the generator learns a physically grounded latent space, while the controller learns to execute latent-conditioned behaviors with dynamical integrity.Extensive experiments on the Unitree G1 humanoid show that our bidirectional optimization improves tracking accuracy and success rates. Across IsaacLab and MuJoCo, the implicit latent-driven pipeline consistently outperforms conventional explicit retargeting baselines in both precision and stability. By coupling diffusion-based motion generation with physical plausibility optimization, our framework provides a practical path toward deployable text-guided humanoid intelligence.

Abstract

Human Motion Generation

Explore similar work

RoboForge: Physically Optimized Text-guided Whole-Body Locomotion for Humanoids

MoLingo: Motion-Language Alignment for Text-to-Human Motion Generation

TEXEDO : Test Time Scaling for Controller-aware Language-conditioned Humanoid Motion Generation