MASkillBlender: Decentralized Whole-Body Coordination for Multi-Humanoid Loco-Manipulation via Skill Blending

Yifan Hu; Luhang Hong; Mingkang Long; Danning Wang; Chengfeng Jia; Rong Su; Junjie Fu; Guanghui Wen

cs.ROOct 1, 2026

MASkillBlender: Decentralized Whole-Body Coordination for Multi-Humanoid Loco-Manipulation via Skill Blending

Authors: Yifan Hu, Luhang Hong, Mingkang Long, Danning Wang, Chengfeng Jia, Rong Su, Junjie Fu, Guanghui Wen

Organizations: Nanyang Technological University · Southeast University · Purple Mountain Laboratories

Abstract

Coordinated multi-humanoid loco-manipulation is promising yet challenging due to high-dimensional whole-body control, decentralized decision making, and scalability. While recent reinforcement learning methods have improved single-humanoid whole-body control, extending them to the multi-humanoid setting remains nontrivial and often requires substantial reward engineering or task-specific design. We propose MASkillBlender, a general multi-agent reinforcement learning framework to achieve decentralized multi-humanoid whole-body coordination. By learning a shared decentralized high-level policy over reusable pre-trained single-humanoid skills, MASkillBlender enables coordinated behaviors using only task-level rewards, without requiring task-specific motion references. To improve learning efficiency, we further introduce a permutation-based data augmentation strategy for homogeneous multi-humanoid systems, and theoretically show that the permuted samples preserve the policy-gradient direction of the original samples under the homogeneous Markov game formulation. We evaluate MASkillBlender on multiple multi-humanoid coordination tasks across two humanoid embodiments. Simulation results demonstrate that the proposed framework consistently achieves strong task performance and enables coordinated behaviors across different tasks and humanoid embodiments.

Figures & tables

Figure 1: MASkillBlender enables decentralized multi-humanoid whole-body coordination across diverse tasks and humanoid embodiments without requiring task-specific motion references.

Figure 2: Overview of MASkillBlender. (a) MASkillBlender enables diverse multi-humanoid coordination across different humanoid embodiments without task-specific motion references or extensive reward engineering. (b)-(c) For each task, relevant pre-trained primitive skills are selected and blended by a shared decentralized high-level policy. (d) The high-level policy is trained with MAPPO using permutation-based data augmentation.

Figure 3: Training curves of all methods on three Unitree H1 tasks.

Task	Metric	MASkillBlender	SkillBlender	TeamHOI
Carry	Return ( $\uparrow$ )	67.928 $\pm$ 1.563	65.855 $\pm$ 8.343	67.870 $\pm$ 9.935
	Error box ( $\downarrow$ )	0.016 $\pm$ 0.008	0.025 $\pm$ 0.050	0.029 $\pm$ 0.102
	Success ( $\uparrow$ )	100.0%	98.0%	96.0%
Push	Return ( $\uparrow$ )	61.137 $\pm$ 5.400	57.550 $\pm$ 5.011	57.991 $\pm$ 5.024
	Error box ( $\downarrow$ )	0.116 $\pm$ 0.063	0.138 $\pm$ 0.069	0.136 $\pm$ 0.075
	Success ( $\uparrow$ )	86.0%	84.0%	78.0%

Table 1: Quantitative comparison of MASkillBlender with SkillBlender and TeamHOI on the Unitree H1 tasks.

Figure 4: Qualitative comparison of MASkillBlender and TeamHOI on the Unitree H1 Carry task. Representative key frames from both methods are shown to illustrate the whole-body coordination behaviors during box lifting and transportation.

Appendix figures & tables11 assets

Supplementary material from the paper’s appendix.

Appendix

Task	Number of Humanoids	Number of Reward Terms	Selected Primitive Skills
			Walking	Reaching	Squatting
Carry	2	2	$\checkmark$	$\checkmark$	$\checkmark$
Push	2	2	$\checkmark$	$\checkmark$
Move	3	3	$\checkmark$

Appendix

Table 2: Task settings.

Task	Actor hidden dimensions	Critic hidden dimensions
Carry	[512, 256, 128]	[768, 256, 128]
Push	[512, 256, 128]	[1024, 512, 128]
Move	[512, 256, 128]	[768, 256, 128]

Appendix

Table 3: Network hyperparameters of MASkillBlender for three Unitree H1 tasks.

Task	Metric	MASkillBlender	MASkillBlender w/o Aug.
Carry	Return ( $\uparrow$ )	67.928 $\pm$ 1.563	45.791 $\pm$ 5.684
	Error box ( $\downarrow$ )	0.016 $\pm$ 0.008	0.179 $\pm$ 0.068
	Success ( $\uparrow$ )	100.0%	6.0%
Push	Return ( $\uparrow$ )	61.137 $\pm$ 5.400	57.130 $\pm$ 9.322
	Error box ( $\downarrow$ )	0.116 $\pm$ 0.063	0.147 $\pm$ 0.098
	Success ( $\uparrow$ )	86.0%	76.0%

Appendix

Table 4: Ablation study of permutation-based data augmentation on the Unitree H1 tasks.

Figure 5: Primitive skill ablation on the Unitree H1. Removing Squatting from Carry or Walking from Push substantially degrades learning performance.

Task	Metric	Nominal	Constrained
Carry	Return ( $\uparrow$ )	67.928 $\pm$ 1.563	63.462 $\pm$ 11.406
	Error box ( $\downarrow$ )	0.016 $\pm$ 0.008	0.073 $\pm$ 0.201
	Success ( $\uparrow$ )	100.0%	92.0%
Push	Return ( $\uparrow$ )	61.137 $\pm$ 5.400	52.744 $\pm$ 12.604
	Error box ( $\downarrow$ )	0.116 $\pm$ 0.063	0.158 $\pm$ 0.089
	Success ( $\uparrow$ )	86.0%	74.0%

Appendix

Table 5: Robustness evaluation of MASkillBlender under deployment noise and delays on the Unitree H1 tasks.

Figure 6: Long-horizon deployment from asymmetric initial configurations on the Unitree H1. The top row shows Carry and the bottom row shows Push . From left to right, the snapshots illustrate the asymmetric initial configuration, one humanoid arriving while the other is still approaching, both humanoids reaching the predefined task poses, and the final coordinated task execution.

Panel of Figure 7 — Figure 7: Sim2Sim deployment of MASkillBlender in MuJoCo on (a) Carry and (b) Move .

Task	Metric	MuJoCo
Carry	Return ( $\uparrow$ )	66.787 $\pm$ 2.221
	Error box ( $\downarrow$ )	0.026 $\pm$ 0.015
Move	Return ( $\uparrow$ )	86.323 $\pm$ 1.430
	Error pos ( $\downarrow$ )	0.013 $\pm$ 0.005
	Error head ( $\downarrow$ )	38.151 $\pm$ 30.906

Appendix

Table 6: Sim2Sim evaluation of MASkillBlender in MuJoCo on the Unitree H1 tasks.

Figure 8: Two-humanoid box-exchange task on the Unitree H1. The snapshots illustrate the execution sequence from the initial configuration to the final box exchange. From (a) to (f), the two humanoids first move their assigned boxes to the exchange region and then deliver the exchanged boxes to their final target positions.

Figure 9: Training curves of MASkillBlender on three Unitree G1 tasks.

Task	Metric	MASkillBlender
Carry	Return ( $\uparrow$ )	91.366 $\pm$ 5.339
	Error box ( $\downarrow$ )	0.025 $\pm$ 0.043
Push	Return ( $\uparrow$ )	68.851 $\pm$ 13.408
	Error box ( $\downarrow$ )	0.113 $\pm$ 0.108
Move	Return ( $\uparrow$ )	81.853 $\pm$ 2.115
	Error pos ( $\downarrow$ )	0.039 $\pm$ 0.006

Appendix

Table 7: Performance of MASkillBlender on the Unitree G1 tasks.

Explore similar work

SkillX: Unified Multi-Skill Policy Learning for Humanoid Soccer
Date pendingZhangchen Ye, Enxuan Ruan, Yifei Bao +10Badminton Human-To-Robot Transfer
Noetix Robotics · Tsinghua University
Distilling Collaborative Dynamics into Latent Space for Implicit Coordination in Decentralized Multi-Agent Manipulation
Jun 22, 2026Chanyoung Park, Minsung Yoon, Andrew Jeong +1Multi-Agent Coordination Global Coordination
School of Computing at the Korea Advanced Institute of Science and Technology (KAIST), Daejeon, 34141, Republic of Korea
Learning Multi-Humanoid Pickup and Transport via Decentralized Object-Centric Control
Sep 15, 2026Bikram Pandit, Mohitvishnu S. Gadde, Aayam Kumar Shrestha +1Bimanual Manipulation Real-World Manipulation Tasks
Dynamics Robotics and AI Lab (DRAIL), Oregon State University, Corvallis, OR, USA

Date pending, cs.ROJ/K move · Enter open · S save

SkillX: Unified Multi-Skill Policy Learning for Humanoid Soccer

Zhangchen Ye, Enxuan Ruan, Yifei Bao, Runhan Huang+9

Noetix Robotics · Tsinghua University

Badminton Human-To-Robot Transfer Whole-Body Control Skills

Open paper AskPDF arXiv

Humanoid soccer is a challenging testbed for dynamic whole-body control, requiring robots to coordinate balance, locomotion, object interaction, and skill switching over long horizons. Existing humanoid sports methods often rely on task-specific multi-stage pipelines, making it difficult to jointly learn and compose multiple object-interactive skills within a single deployable policy. To address this, we present SkillX, a unified reinforcement learning framework that learns and composes multiple atomic soccer skills through a single command-conditioned policy. SkillX integrates three core designs: skill-specific adversarial motion priors, skill-specific critics, and an object-aware temporal encoder, enabling the robot to execute atomic skills and transition among them such as dribbling, trapping, and shooting. Experiments in simulation and on a real Noetix E1 humanoid demonstrate robust multi-skill execution, long-horizon skill composition, and successful sim-to-real deployment.

Abstract

Explore similar work

SkillX: Unified Multi-Skill Policy Learning for Humanoid Soccer

Distilling Collaborative Dynamics into Latent Space for Implicit Coordination in Decentralized Multi-Agent Manipulation

Learning Multi-Humanoid Pickup and Transport via Decentralized Object-Centric Control