OmniXtreme: Breaking the Generality Barrier in High-Dynamic Humanoid Control

Yunshen Wang; Shaohang Zhu; Peiyuan Zhi; Yuhan Li; Jiaxin Li; Yong-Lu Li; Yuchen Xiao; Xingxing Wang

cs.ROFeb 27, 2026

OmniXtreme: Breaking the Generality Barrier in High-Dynamic Humanoid Control

Authors: Yunshen Wang, Shaohang Zhu, Peiyuan Zhi, Yuhan Li, Jiaxin Li, Yong-Lu Li, Yuchen Xiao, Xingxing Wang, +2 more

Organizations: State Key Laboratory of General Artificial Intelligence, Beijing Institute for General Artificial Intelligence (BIGAI) · Joint Laboratory of Embodied AI and Humanoid Robots, BIGAI & Unitree Robotics · Shanghai Jiao Tong University · University of Science and Technology of China · Huazhong University of Science and Technology · Beijing Institute of Technology · Unitree Robotics

Abstract

High-fidelity motion tracking serves as the ultimate litmus test for generalizable, human-level motor skills. However, current policies often hit a "generality barrier": as motion libraries scale in diversity, tracking fidelity inevitably collapses - especially for real-world deployment of high-dynamic motions. We identify this failure as the result of two compounding factors: the learning bottleneck in scaling multi-motion optimization and the physical executability constraints that arise in real-world actuation. To overcome these challenges, we introduce OmniXtreme, a scalable framework that decouples general motor skill learning from sim-to-real physical skill refinement. Our approach uses a flow-matching policy with high-capacity architectures to scale representation capacity without interference-intensive multi-motion RL optimization, followed by an actuation-aware refinement phase that ensures robust performance on physical hardware. Extensive experiments demonstrate that OmniXtreme maintains high-fidelity tracking across diverse, high-difficulty datasets. On real robots, the unified policy successfully executes multiple extreme motions, effectively breaking the long-standing fidelity-scalability trade-off in high-dynamic humanoid control.

Figures & tables

Fig. 2 : Overview of the OmniXtreme . (a) Pretraining phase: A unified base policy is trained via DAgger-based Flow Matching to aggregate diverse motion priors from different motion tracking experts. (b) Post-training phase: The base policy is frozen while a residual policy is optimized under stringent motor constraints, extensive domain randomization, and power-safety regularization to bridge the sim-to-real gap. (c) Onboard deployment: The whole inference pipeline is real-time and executed entirely onboard, facilitating robust and agile control in physical environments.

Parameter Item	Moderate	Aggressive
Noise and Domain Randomization
Joint Position (rad)	$\pm 0.01$	$\pm 0.01$
Joint Velocity (rad/s)	$\pm 0.5$	$\pm 0.5$
Angular Velocity (rad/s)	$\pm 0.2$	$\pm 0.2$
Torso 6D Rotation (rad)	$\pm 0.05$	$\pm 0.05$
Base CoM Offset (m)	$x$ : $\pm 0.025,y,z$ : $\pm 0.05$	$x$ : $\pm 0.025,y,z$ : $\pm 0.05$

TABLE I : Configurations for noise, domain randomization, and termination thresholds during pre-training and post-training phases. Here

\pm x

denotes

[-x,x]

Method	LaFAN1+XtremeMotion				XtremeMotion				Unseen Motions
	MPJPE $\downarrow$	$\Delta$ vel $\downarrow$	$\Delta$ acc $\downarrow$	Succ.(%) $\uparrow$	MPJPE $\downarrow$	$\Delta$ vel $\downarrow$	$\Delta$ acc $\downarrow$	Succ.(%) $\uparrow$	MPJPE $\downarrow$	Succ.(%) $\uparrow$
From-scratch RL [ 28 ]	47.95	10.03	3.27	82.95	54.19	14.04	4.04	79.45	56.87	85.29
Specialist $\rightarrow$ Unified MLP [ 58 ]	33.35	6.70	2.11	94.91	43.43	11.38	2.51	89.22	58.94	85.95
OmniXtreme (Pretrain only)	32.65	6.34	2.04	97.17	37.11	10.46	2.39	95.16	56.25	89.23
OmniXtreme (Pretrain + Post-train)	30.93	6.19	2.13	98.54	36.17	9.94	2.58	95.64	56.05	89.54

TABLE II: Scalable high-fidelity motion tracking under diverse motion sets. OmniXtreme consistently achieves lower kinematic errors and higher success rates than baselines, particularly on high-dynamic and unseen motions.

Skill	#Motions	Attempts	Success (%) $\uparrow$
Flip	7	55	96.36
Handspring	5	35	88.57
Acrobatics	4	15	80.00
Breakdance	5	22	86.36
Martial arts	3	30	93.33
Total	24	157	91.08

TABLE III: Real-world evaluation of OmniXtreme on Unitree G1. We evaluate OmniXtreme on physical hardware using motions drawn from the XtremeMotion motion library.

Fig. 3 : Fidelity–scalability trade-off. Tracking success rate as we progressively scale motion diversity and difficulty, while evaluating all policies on a fixed set of the first 10 motions.

Fig. 4 : Capacity scaling. Tracking fidelity and robustness as a function of model capacity. OmniXtreme benefit more strongly from scaling, while conventional MLP controllers saturate earlier.

Skill	None	+MC	+MC+ADR	Full (+MC+ADR+PS)
Flip	$\triangle$	$\checkmark$	$\checkmark$	$\checkmark$
Breakdance	$\triangle$	$\triangle$	$\checkmark$	$\checkmark$
Acrobatics	$\times$	$\triangle$	$\circleddash$	$\checkmark$

TABLE IV: Ablation of post-training mechanisms. Real-world executability of different skills under incremental post-training mechanisms. None : base pretrained policy only; MC : motor constraints; ADR : aggressive domain randomization; PS : power-safety regularization (overcurrent / regenerative protection).

\checkmark

: stable execution;

\triangle

: unstable or inconsistent execution;

\times

: consistent failure;

\circleddash

: failures primarily associated with power-safety protection, such as overcurrent or excessive regenerative braking.

Fig. 5 : Qualitative results. Representative real-world rollouts produced by OmniXtreme , executing qualitatively distinct whole-body motions across diverse styles and contact patterns, including flips, acrobatics, breakdance, and martial-arts behaviors. The results illustrate stable and coordinated execution under rapid contact transitions and timing-sensitive phases on physical hardware.

Motion ID	Motion Source	Motion Description
1	CMU-85-05	Handstand walk.
2	CMU-85-10	Handstand spin.
3	CMU-88-09	Back Handspring with a full twist.
4	CMU-88-08	Back Handspring with a half twist.
5	CMU-90-08	Aerial Cartwheel
6	CMU-85-14	B-boying with a rapid-fire string of back handsprings.

TABLE V : Motion information summary.

Architecture
Hyperparameter	Value
Actor MLP hidden dimensions	[512, 256, 128]
Critic MLP hidden dimensions	[512, 256, 128]
Activation function	ELU
Training
Steps per environment	24

TABLE VI : Motion tracking hyperparameters for teacher training [ 28 ] .

Reward Term	Weight	Expression
Global Torso Position	0.5	$\exp(-\\|p_{err}\\|^{2}/0.3^{2})$
Global Torso Orientation	0.5	$\exp(-\\|\theta_{err}\\|^{2}/0.4^{2})$
Relative Body Position	1.0	$\exp(-\\|p_{rel\_err}\\|^{2}/0.3^{2})$
Relative Body Orientation	1.0	$\exp(-\\|\theta_{rel\_err}\\|^{2}/0.4^{2})$
Body Linear Velocity	1.0	$\exp(-\\|v_{err}\\|^{2}/1.0^{2})$
Body Angular Velocity	1.0	$\exp(-\\|\omega_{err}\\|^{2}/3.14^{2})$

TABLE VII : Reward function terms and expressions for Teacher Policy training [ 28 ] .

Reward Term	Weight	Expression
Global Torso Position	0.5	$\exp(-\\|p_{err}\\|^{2}/0.3^{2})$
Global Torso Orientation	0.5	$\exp(-\\|(\theta_{err}\\|^{2}/0.4^{2})$
Relative Body Position	1.0	$\exp(-\\|p_{rel\_err}\\|^{2}/0.3^{2})$
Relative Body Orientation	1.0	$\exp(-\\|\theta_{rel\_err}\\|^{2}/0.4^{2})$
Body Linear Velocity	1.0	$\exp(-\\|v_{err}\\|^{2}/1.0^{2})$
Body Angular Velocity	1.0	$\exp(-\\|\omega_{err}\\|^{2}/3.14^{2})$

TABLE VIII : Reward function terms and expressions for Residual Policy training. Refer to Sec. V-K for further details regarding power safety regularization.

Actuator	$\tau_{y1}$	$\tau_{y2}$	$v_{x1}$	$v_{x2}$	$\mu_{s}$	$v_{act}$	$\mu_{d}$	$I$
5020-16	24.8	31.9	30.86	40.13	0.6	0.01	0.06	3.610e-03
7520-14.3	71.0	83.3	22.63	35.52	1.6	0.01	0.16	1.018e-02
7520-22.5	111.0	131.0	14.5	22.7	2.4	0.01	0.24	2.510e-02
4010-25	4.8	8.6	15.3	24.76	0.6	0.01	0.06	4.250e-03

TABLE IX : Actuator modeling parameters.

Joint Name	Motor Model
Hip Pitch Joint	7520-22.5
Hip Roll Joint	7520-22.5
Knee Joint	7520-22.5
Hip Yaw Joint	7520-14.3
Ankle Pitch Joint	5020
Ankle Roll Joint	5020

TABLE X : Unitree G1 Joint-to-Motor Mapping.

Skill category	Motion IDs
Flip	5, 9, 10, 15, 16, 18, 22
Handspring	2, 3, 4, 19, 23
Acrobatics	1, 6, 7, 20, 21
Breakdance	11, 12, 13, 14
Martial arts	8, 17, 38

TABLE XI : Skill-level grouping of real-world evaluation motions. Each skill category groups multiple motion instances with similar semantic meaning and dynamic structure. Motion IDs correspond to retargeted motions used during hardware evaluation.

Training motions	Motion IDs
10	3–10, 13, 14
20	2–10, 13-22
50	All

TABLE XII : Motion subsets used in Q2 fidelity–scalability analysis. All evaluations are performed on the same first 10 extreme motions. Larger training sets (20 and all motions) extend this core set with additional diverse extreme motions.

Explore similar work

Extreme-RGMT: Continual Learning of Highly Dynamic Skills for Robust Generalist Humanoid Control
Jul 22, 2026Yubiao Ma, Han Yu, Kai Guo +5Continual Robot Learning Robot Skill Learning
School of Automation, Beijing Institute of Technology, Beijing 100081, China. · Humanoid Robotics (Shanghai) Co., Ltd., Shanghai 201203, China. · School of Mechanical Engineering, Shandong University, Jinan 250061, China.
OmniTrack: General Motion Tracking via Physics-Consistent Reference
Feb 27, 2026Yuhan Li, Peiyuan Zhi, Yunshen Wang +6Humanoid Motion Tracking Humanoid Robot Control
Huazhong University of Science and Technology · State Key Lab of General AI, Beijing Institute for General Artificial Intelligence (BIGAI) · Shanghai Jiao Tong University
Iterative Closed-Loop Motion Synthesis for Scaling the Capabilities of Humanoid Control
Feb 25, 2026Weisheng Xu, Qiwei Wu, Jiaxi Zhang +7Humanoid Robot Control Robot Motion Generation
Hong Kong University of Science and Technology (Guangzhou) · University of Oxford

Jul 22, 2026J/K move · Enter open · S save

Extreme-RGMT: Continual Learning of Highly Dynamic Skills for Robust Generalist Humanoid Control

Yubiao Ma, Han Yu, Kai Guo, Changtai Lv+4

School of Automation, Beijing Institute of Technology, Beijing 100081, China. · Humanoid Robotics (Shanghai) Co., Ltd., Shanghai 201203, China. · School of Mechanical Engineering, Shandong University, Jinan 250061, China.

Continual Robot Learning Robot Skill Learning Humanoid Robot Control Robot Policy Generalization Continual RL

Open paper AskPDF arXiv

Humans can progressively acquire highly dynamic motor skills while preserving reliable everyday motor abilities. In contrast, existing humanoid controllers face a trade-off between generalist and specialist capabilities: generalist motion tracking policies struggle to reliably execute rare highly dynamic motions, whereas specialist training can degrade previously acquired behaviors. We introduce Extreme-RGMT, a two-stage continual learning framework for robust generalist humanoid control. The method first learns a generalist motion-tracking base policy from diverse multi-source motion data, then employs an asymmetric skill acquisition and capability consolidation mechanism to constrain policy drift on mastered motions while emphasizing difficult dynamic segments. To address the scarcity of highly dynamic motions, their high failure rates, and the resulting shortage of informative samples, Extreme-RGMT combines difficulty-aware sampling with advantage-prioritized trajectory resampling to emphasize critical segments. Experiments show that Extreme-RGMT achieves state-of-the-art generalist whole-body motion-tracking performance, including substantially improved completion of challenging highly dynamic motions. The resulting controller directly executes diverse unseen highly dynamic motions under fixed references and online inertial motion-capture inputs, advancing generalist whole-body motion-tracking controllers toward highly dynamic motor capabilities at the human-expert level.

Abstract

Explore similar work

Extreme-RGMT: Continual Learning of Highly Dynamic Skills for Robust Generalist Humanoid Control

OmniTrack: General Motion Tracking via Physics-Consistent Reference

Iterative Closed-Loop Motion Synthesis for Scaling the Capabilities of Humanoid Control