Humanoid Badminton: Learning Dynamic Racket Skills from Limited Human Motion Data

Jingzhi Cui; Zhexiong Wang; Bangjie Xu; Pengyu Zhao; Youyuan Li; Zhi Su; Peng Ren; Mengdi Xu

cs.ROSep 25, 2026

Humanoid Badminton: Learning Dynamic Racket Skills from Limited Human Motion Data

Authors: Jingzhi Cui, Zhexiong Wang, Bangjie Xu, Pengyu Zhao, Youyuan Li, Zhi Su, Peng Ren, Mengdi Xu, +4 more

Organizations: Tsinghua University · Hong Kong Embodied AI Lab · Zhejiang University · Beijing University of Civil Engineering and Architecture · DeepCybo · The Chinese University of Hong Kong

Abstract

High-speed racket sports provide a demanding testbed for humanoid robots, requiring time-critical decisions, precise striking, and dynamic whole-body coordination. In badminton, fast-changing shuttle trajectories require timely contact decisions, while successful returns demand precise racket pose and velocity within a brief contact window and across a broad three-dimensional striking workspace. Human motion data provide valuable priors for such athletic skills, but usable badminton references are limited and imperfect. Direct tracking provides insufficient executable variation for diverse shuttle conditions, while purely task-driven optimization may produce unnatural motion. To address these challenges, we present a three-stage hierarchical reinforcement learning framework for dynamic humanoid badminton. First, task-randomized motion augmentation expands sparse annotated hitting events into executable target-conditioned stroke variations, forming a continuous latent skill space. Second, a high-level planner outputs continuous latent skill codes to compose these skills online according to the observed shuttle state. Third, a context-conditioned adversarial regularizer encourages more natural planner-level skill usage while preserving return performance. When deployed on a real humanoid robot, our system achieves sustained multi-skill rallies with human players, including forehand, backhand, and highly dynamic jump returns. This is the first real-world humanoid racket-sport system to demonstrate multi-skill human--robot rallies including highly dynamic jump returns.

Figures & tables

Figure 1: Real-world athletic badminton with a humanoid robot. (a) The robot rallies with a human player in the real world, requiring timely hitting decisions, coordinated footwork and racket control, and whole-body balance. (b) A close-up of a highly dynamic jump return, which couples rapid takeoff, aerial striking with a high-speed racket swing, and stable landing.

Figure 2: Overview of the proposed hierarchical learning framework. We first reconstruct and retarget human badminton videos, annotate hitting events, and perform task-randomized motion augmentation by sampling racket targets for each annotated hitting event, including contact position $p_{\mathrm{target}}$ , racket velocity $v_{\mathrm{target}}$ , and racket-face direction $d_{\mathrm{target}}$ . In (a) , a target-conditioned skill encoder maps reference-motion information, the sampled target, hit timing, and robot context to a latent skill code $z_{t}$ , while a low-level controller $\pi_{\phi}$ maps the latent code and robot context to joint actions. In (b) , the low-level controller is frozen, and a high-level planner $\pi_{\theta}$ observes shuttle history and robot state information to output continuous latent codes for shuttle returns. In (c) , a context-conditioned discriminator distinguishes Stage 1-generated latent–context pairs from planner-generated pairs, providing a latent-style reward for regularizing planner-level skill usage.

Method	SR (%) $\uparrow$	SR-F (%)	SR-B (%)	SR-J (%)	AC $\uparrow$	Joint Acc. $\downarrow$	Joint Jerk $\downarrow$	Torque $\downarrow$	JFID $\downarrow$
Direct PPO	79.0	0.0	25.1	53.9	4.14	59.63	26743.64	8.09	1226.88
AMP	84.6	38.4	0.0	46.2	5.06	35.48	13420.18	6.55	295.96
Ours	88.3	34.6	32.9	20.8	5.98	34.28	19318.24	6.13	42.61

Table 1: Comparison with baseline methods. SR-F, SR-B, and SR-J denote the percentages of all served shuttles successfully returned using forehand, backhand, and jump-return behaviors, respectively. Lower Joint Acceleration, Joint Jerk, Joint Torque, and JFID indicate better motion-quality or distributional metrics.

Method	SR (%) $\uparrow$	SR-F (%)	SR-B (%)	SR-J (%)	AC $\uparrow$
w/o Task-Rand. Aug.	57.0	40.2	0.0	16.8	2.38
Stage 2 Only	89.2	24.5	25.6	39.1	5.92

Table 2: Ablation on task-randomized motion augmentation. SR-F, SR-B, and SR-J decompose the overall success rate into forehand, backhand, and jump-return behaviors, respectively.

Figure 3: Left: Randomized shuttle initialization generates diverse incoming trajectories, including high trajectories that create opportunities for jump returns. Right: t-SNE visualization of target-conditioned latent codes collected from Stage 1 rollouts under a fixed reference motion and hitting event. Color indicates the remaining time to hit. Randomized-target codes spread around the corresponding nominal-target codes, forming local neighborhoods of executable stroke variations.

Method	SR (%) $\uparrow$	SR-F (%)	SR-B (%)	SR-J (%)	AC $\uparrow$
w/o Staged Training	55.9	55.1	0.0	0.8	2.42
Ours	88.3	34.6	32.9	20.8	5.98

Table 3: Ablation on the staged training pipeline.

Method	SR (%) $\uparrow$	SR-F (%)	SR-B (%)	SR-J (%)	AC $\uparrow$	Joint Acc. $\downarrow$	Joint Jerk $\downarrow$	Torque $\downarrow$	JFID $\downarrow$
Stage 2 Only	89.2	24.5	25.6	39.1	5.92	34.62	23125.80	7.05	205.65
Ours	88.3	34.6	32.9	20.8	5.98	34.28	19318.24	6.13	42.61

Table 4: Effect of planner-level latent skill regularization. JFID is computed with respect to the Stage 1 positive rollout distribution.

Setting	SR (%) $\uparrow$	SR-F (%)	SR-B (%)	SR-J (%)	AC $\uparrow$	Max CH $\uparrow$
Real World	89.4	60.2	14.2	15.0	8.42	23

Table 5: Real-world deployment results. SR-F, SR-B, and SR-J denote successful forehand, backhand, and jump returns over all robot attempts. Consecutive-hit statistics are uncapped.

Figure 4: Representative real-world executions of the deployed humanoid badminton policy. The robot performs forehand, backhand, and jump returns with coordinated whole-body movement for shuttle interception.

Appendix figures & tables2 assets

Supplementary material from the paper’s appendix.

Appendix

Type	Description
Pre-contact	Encourages lateral alignment with the incoming shuttle, reduces the racket–shuttle distance, and aligns the pre-contact racket velocity with a feasible hitting direction.
Impact	Encourages sufficient racket velocity and racket-face alignment with the desired return direction at first racket–shuttle contact.
Post-contact	Encourages the returned shuttle to travel toward the opponent side with sufficient forward velocity.
Jump return	Encourages elevated contact, upward whole-body motion, and aerial striking for suitable high shuttles, while penalizing unnecessary jumps.
Regularization	Encourages upright body orientation and sufficient body height, while penalizing prolonged racket–shuttle contact, abrupt action changes, joint-limit violations, self-collisions, and overly rapid stepping.

Appendix

Table 6: Summary of reward groups used for high-level planner training.

Figure 5: Planner latent-space evolution. PCA visualization of successful forehand, backhand, and jump-return latent codes at (a)

400

ms before contact, (b)

200

ms before contact, and (c) the contact frame, using a shared PCA basis. The return types progressively separate as contact approaches. The trajectory in (c) shows a consecutive three-return sequence covering all three return types; numbered markers indicate the return order and type.

Explore similar work

Humanoid Whole-Body Badminton via an Annealed Reinforcement Learning Curriculum
Nov 14, 2025Chenhao Liu, Leyun Jiang, Ningyuan Tian +4RL for Robotics Reinforcement Learning
Learning Human-Like Badminton Skills for Humanoid Robots
Feb 9, 2026Yeke Chen, Shihao Dong, Xiaoyu Ji +12Humanoid Robot Control Robotic RL
Adaptive Robotic Controls Lab (ArcLab), Department of Mechanical Engineering, The University of Hong Kong, Hong Kong SAR, China · EngineAI Robotics Technology Co., Ltd., Shenzhen, China
Learning Athletic Humanoid Tennis Skills from Imperfect Human Motion Data
Mar 13, 2026Zhikai Zhang, Haofei Lu, Yunrui Lian +12Robotics Humanoid Robot Control
Tsinghua University · Galbot Inc. · Shanghai AI Laboratory +2

Nov 14, 2025J/K move · Enter open · S save

Humanoid Whole-Body Badminton via an Annealed Reinforcement Learning Curriculum

Chenhao Liu, Leyun Jiang, Ningyuan Tian, Yibo Wang+3

RL for Robotics Reinforcement Learning Humanoid Robot Control Curriculum RL Robotics

Open paper AskPDF arXiv

Humanoid robots have demonstrated strong capabilities for interacting with static scenes across locomotion and manipulation, yet dynamic real-world interactions remain challenging. As a step toward fast-moving object interactions, we present an RL training pipeline that yields a unified whole-body controller for humanoid badminton, coordinating footwork and striking without motion priors or expert demonstrations. In badminton, locomotion and striking are tightly coupled, making the final hitting objective difficult to optimize directly due to sparse rewards and conflicting gradients. We address this with an annealed curriculum that first stabilizes learning through auxiliary locomotion objectives, then progressively removes them to focus optimization on the final hitting objective. For deployment, we use an Extended Kalman Filter (EKF) to estimate and predict shuttlecock trajectories for target striking, and also develop a prediction-free variant that removes the EKF and explicit prediction. We validate the framework in simulation and on hardware. In simulation, two robots sustain a rally of 21 consecutive hits. In real-world tests with both machine-fed shuttles and human-robot rallies, the robot achieves outgoing shuttle speeds up to 19.1 m/s. Moreover, the prediction-free variant attains comparable performance to the EKF-based policy. Overall, our approach enables dynamic yet precise goal striking in humanoid badminton and suggests a path toward more dynamics-critical whole-body interaction tasks.

Abstract

Explore similar work

Humanoid Whole-Body Badminton via an Annealed Reinforcement Learning Curriculum

Learning Human-Like Badminton Skills for Humanoid Robots

Learning Athletic Humanoid Tennis Skills from Imperfect Human Motion Data