Dynamic Whole-Body Dancing with Humanoid Robots -- A Model-Based Control Approach

Shibowen Zhang; Jiayang Wu; Guannan Liu; Helin Zhu; Junjie Liu; Zhehan Li; Junhong Guo; Xiaokun Leng

cs.ROApr 5, 2026

Dynamic Whole-Body Dancing with Humanoid Robots -- A Model-Based Control Approach

Authors: Shibowen Zhang, Jiayang Wu, Guannan Liu, Helin Zhu, Junjie Liu, Zhehan Li, Junhong Guo, Xiaokun Leng,

Organizations: Department of Automation, University of Science and Technology of China, Hefei 230027, China. · State Key Laboratory of General Artificial Intelligence, Beijing Institute for General Artificial Intelligence (BIGAI), Beijing 100080, China · Department of Automation, Tsinghua University, Beijing 100084, China. · School of Artificial Intelligence, Xidian University, Xi’an 710126, China. ·

Abstract

This paper presents an integrated model-based framework for generating and executing dynamic whole-body dance motions on humanoid robots. The framework operates in two stages: offline motion generation and online motion execution, both leveraging future state prediction to enable robust and dynamic dance motions in real-world environments. In the offline motion generation stage, human dance demonstrations are captured via a motion capture (MoCap) system, retargeted to the robot by solving a Quadratic Programming (QP) problem, and further refined using Trajectory Optimization (TO) to ensure dynamic feasibility. In the online motion execution stage, a centroidal dynamics-based Model Predictive Control (MPC) framework tracks the planned motions in real time and proactively adjusts swing foot placement to adapt to real world disturbances. We validate our framework on the full-size humanoid robot Kuavo 4Pro, demonstrating the dynamic dance motions both in simulation and in a four-minute live public performance with a team of four robots. Experimental results show that longer prediction horizons improve both motion expressiveness in planning and stability in execution.

Figures & tables

Fig. 1 : Dynamic whole-body dance motions performed by a team of four humanoid robots.

Fig. 2 : Overview of the dance motion generation framework. (a) The mocap system captures the motion patterns of markers on the human demonstrator using a high-resolution camera group. (b) Geometric motion retargeting morphs the human motions into robot motions by minimizing the position error between robot joints and human skeleton joints. The corresponding joints of these two models are marked in the same color in (b) and (c). (c) A snapshot of the simulation using the trajectory output by geometric motion retargeting shows a dynamically infeasible motion. (d) Dynamic motion retargeting adapts the trajectory to meet dynamic and kinematic constraints via to . (e) A snapshot of the simulation using the trajectory after adjustment by dynamic motion retargeting shows a dynamically feasible motion.

Fig. 3 : The placement of the markers on the human body. The markers are grouped by different colors, red for head, green for torso, yellow for arms, pink for waist, purple for legs, and blue for feet.

Fig. 4 : Overview of the online motion execution framework. (i) The centroidal mpc takes the offline-generated dance trajectory and the robot’s current state as inputs to compute an optimal reference state trajectory. (ii) The whole-body control tracks the reference state trajectory and generates torque control commands for the motors.

Tasks	Constraints
1) Stance Leg Joint Motion Tracking	1) Joint Limits
2) Swing Leg Joint Motion Tracking	2) Friction Cone
3) Base Angular Motion Tracking	3) Robot Dynamics
4) Arm Joint Motion Tracking
5) com Tracking
6) Contact Force Tracking

TABLE I : Overview of the whole-body control tasks and constraints

Fig. 5 : Hardware design and configuration of the Kuavo 4Pro humanoid robot. The total height of the robot is 1.66 m with a mass of 55 kg. Each leg contains 6 dof and each arm contains 7 dof . The head integrates a 2- dof actuation system with a Realsense D435 camera and a Livox Mid360 lidar. The robot contains two computers: an Intel NUC computer (Core i9-13900H) for control and communications, and an NVIDIA Jetson AGX Orin (64 GB) for vision and AI algorithms.

Fig. 6 : Experimental results on enforcing dynamic constraints in the dynamic retargeting module. (a) Sequential snapshots of the dance motion in simulation. (b) The top plot shows the knee joint position of the motion pointed by the red arrow. The blue and red lines display the joint position generated by geometric and dynamic motion retargeting, respectively. The gray dashed line shows the knee joint limit. The bottom plot is the knee joint torque of the motion generated by dynamic motion retargeting.

Fig. 7 : Experimental results on evaluating the momentum modulation in the dynamic retargeting module. (a) Snapshots of the experimental motion sequence. (b) Angular velocities of the arm and swing leg generated by geometric (blue) and dynamic (red) motion retargeting. (c) Stance foot contact force in the y direction of the motion generated by geometric and dynamic motion retargeting. (d) zmp trajectories corresponding to geometric (left) and dynamic (right) motion retargeting. The green rectangle indicates the support polygon.

Fig. 8 : Experimental results on evaluating the dynamic capability under different prediction horizons. (a) Average velocity of the swing leg. (b) Average velocity of the com .

Fig. 9 : Experimental results for assessing the online motion execution module’s ability to adapt to errors under different prediction horizons. (a) The top snapshot shows the motion with a prediction horizon of 0.6 s, while the bottom snapshot corresponds to a prediction horizon of 1.2 s. (b) Knee joint trajectories for prediction horizons of 0.6 s (blue) and 1.2 s (red). (c) zmp trajectory (blue) and support polygon of adjusted (green) and reference (red) foot placement. The arrow points to the center of the support polygon.

Prediction Horizon (s)	Avg Solve Time (ms)
0.8	14.99
1.0	18.06
1.2	20.64
1.4	24.29

TABLE II : Average solve times of mpc for different prediction horizons.

Fig. 10 : The Kuavo 4Pro humanoid robot performing dynamic whole-body dance motions at the opening ceremony of the 2025 Zhongguancun Forum.

Robot Motion Generation

Robot motion generation creates feasible, task-directed trajectories for robots to move, manipulate objects, and interact with people and environments.

See all 198 papers

KungfuBot: Physics-Based Humanoid Whole-Body Control for Learning Highly-Dynamic Skills
Jun 15, 2025Weiji Xie, Jinrui Han, Jiakun Zheng +6Humanoid Robot Control Robot Imitation Learning
Institute of Artificial Intelligence (TeleAI), China Telecom · Shanghai Jiao Tong University · East China University of Science and Technology +2
Whole-Body Humanoid Control via Stochastic Motion Tokenization
Oct 27, 2026Michael Drolet, Oleg Arenz, Jan PetersHumanoid Robot Control Humanoid Motion Tracking
Technical University of Darmstadt · German Research Center for AI (DFKI) · Robotics Institute Germany (RIG)
Do You Have Freestyle? Expressive Humanoid Locomotion via Audio Control
Dec 29, 2025Zhe Li, Yangyang Wei, Boan Zhu +9Humanoid Robot Locomotion Human Motion Generation
Nanyang Technological University · BAAI · Harbin Institute of Technology +4

Jun 15, 2025J/K move · Enter open · S save

KungfuBot: Physics-Based Humanoid Whole-Body Control for Learning Highly-Dynamic Skills

Weiji Xie, Jinrui Han, Jiakun Zheng, Huanyu Li+5

Institute of Artificial Intelligence (TeleAI), China Telecom · Shanghai Jiao Tong University · East China University of Science and Technology+2

Humanoid Robot Control Robot Imitation Learning Humanoid Motion Tracking

Open paper AskPDF arXiv

Humanoid robots are promising to acquire various skills by imitating human behaviors. However, existing algorithms are only capable of tracking smooth, low-speed human motions, even with delicate reward and curriculum design. This paper presents a physics-based humanoid control framework, aiming to master highly-dynamic human behaviors such as Kungfu and dancing through multi-steps motion processing and adaptive motion tracking. For motion processing, we design a pipeline to extract, filter out, correct, and retarget motions, while ensuring compliance with physical constraints to the maximum extent. For motion imitation, we formulate a bi-level optimization problem to dynamically adjust the tracking accuracy tolerance based on the current tracking error, creating an adaptive curriculum mechanism. We further construct an asymmetric actor-critic framework for policy training. In experiments, we train whole-body control policies to imitate a set of highly-dynamic motions. Our method achieves significantly lower tracking errors than existing approaches and is successfully deployed on the Unitree G1 robot, demonstrating stable and expressive behaviors. The project page is https://kungfubot.github.io.

Abstract

Robot Motion Generation

Explore similar work

KungfuBot: Physics-Based Humanoid Whole-Body Control for Learning Highly-Dynamic Skills

Whole-Body Humanoid Control via Stochastic Motion Tokenization

Do You Have Freestyle? Expressive Humanoid Locomotion via Audio Control