ATHLETE: Learning Reactive Humanoid Tennis via Trajectory-Guided Motion Matching

Zhihao Tian; Zifan Wang; Enzheng Zhu; Xun Yang; Jialong Qin; Mengbo Wang; Junwei Liang

cs.ROOct 9, 2026

ATHLETE: Learning Reactive Humanoid Tennis via Trajectory-Guided Motion Matching

Authors: Zhihao Tian, Zifan Wang, Enzheng Zhu, Xun Yang, Jialong Qin, Mengbo Wang, Junwei Liang

Organizations: The Hong Kong University of Science and Technology (Guangzhou), Guangzhou, China.

Abstract

Humanoid tennis requires reactive whole-body coordination under rapidly changing ball dynamics, yet collecting task-specific motion-capture demonstrations that cover diverse approach states, contact locations, and timing conditions is costly. Inspired by how humans acquire new skills by combining reusable motor experience with task-specific demonstrations, we present ATHLETE, a framework for learning reactive humanoid tennis from generic locomotion data and ordinary tennis videos. ATHLETE combines LAFAN1 locomotion with video-derived striking motions and uses Trajectory-Guided Motion Matching (TGMM) to construct diverse task-conditioned whole-body references. These references are transformed into a reactive policy through physics-based teacher--student learning with online DAgger distillation, followed by PPO fine-tuning for the tennis task. The resulting policy responds directly to robot and ball observations without requiring reference-motion inputs at deployment. On a real humanoid robot, ATHLETE achieves forehand and backhand hit rates of $90.00\%$ and $86.67\%$ , respectively, and enables multi-shot rallies with human players. These results demonstrate that ATHLETE can learn dynamic humanoid tennis skills from reusable locomotion data and video demonstrations. Result videos and code can be found at https://anonymous.4open.science/w/ATHLETE2027/.

Figures & tables

Fig. 2: Strike-point height and XY distribution of source skills and generated references. (a) strike height distribution of source tennis references. (b) XY distribution of strike points in local frames.

Fig. 3: Strike-Motion generation using Trajectory-Guided Motion Matching (TGMM). The yellow ball marks the target point. The blue trajectories are optimized for smoothness under planar motion constraints and then converted into whole-body motion. The translucent blue and green robots depict the locomotion and striking phases, respectively. The green curve traces the trajectory of the racket’s sweet spot.

Fig. 4: Overview of the ATHLETE framework. (1) We construct reference motions from tennis videos and a subset of the LAFAN1 dataset using TGMM, followed by mimic-policy rollouts for dynamic correction. (2) We train a time-adaptive teacher policy to expand the range of incoming ball trajectories that can be matched to the generated references. (3) We distill the teacher into a reactive student policy through online DAgger and privileged representation alignment. (4) We fine-tune the student using PPO with task-specific rewards and domain randomization to improve ball contact and return accuracy. (5) The resulting student policy is deployed on a humanoid robot to enable real-world tennis rallies with human players, using a motion-capture system for ball-state estimation.

Term	Teacher	Critic	Actor
Joint positions / velocities	$\checkmark$	$\checkmark$	$\checkmark$
Previous action	$\checkmark$	$\checkmark$	$\checkmark$
Body angular velocity	$\checkmark$	$\checkmark$	$\checkmark$
Projected gravity	$\times$	$\checkmark$	$\checkmark$
Root linear velocity	$\times$	$\checkmark$	$\times$
World-frame root position	$\times$	$\checkmark$	$\checkmark$

TABLE I: Policy observation

Fig. 5: Spatiotemporal matching success rates of original and TGMM-generated reference libraries with varying numbers of qualified, forward-simulated ball trajectories per group, $K$ . A group is successful if at least one candidate trajectory matches a reference motion.

Policy	Hit rate (%) $\uparrow$		Return rate (%) $\uparrow$		Landing err. (m) $\downarrow$		Sweet-spot dist. (cm) $\downarrow$
	Forehand	Backhand	Forehand	Backhand	Forehand	Backhand	Forehand	Backhand
w/o PPO fine-tuning	1.26	5.67	0.00	0.00	5.465	5.336	14.09	13.92
Full method	94.97	95.04	91.19	92.20	1.150	1.314	5.61	6.54

TABLE II: Paired policy comparison on simulated incoming balls, with 159 forehand-side and 141 backhand-side trials per policy. Hit and return rates use all trials; return denotes net clearance. Landing error is the mean horizontal distance from the first post-strike landing to the target. Sweet-spot distance is the mean 3D distance from the ball center to the racket sweet spot at first contact.

Fig. 6: Teacher contact timing across deadline ratios. Dots show absolute timing errors $|e|=|t_{\mathrm{contact}}-T_{i}|$ for 493 physical racket contacts, where $T_{i}$ and $t_{\mathrm{contact}}$ denote the target and actual contact times. Curves show binned mean absolute errors and 95th percentiles; shading marks errors within 50 ms.

Method	Hit rate (%) $\uparrow$		Return rate (%) $\uparrow$
	Forehand	Backhand	Forehand	Backhand
LATENT [ 1 ]	—	—	90.90	77.78
AdaPT [ 2 ]	52.00	50.67	21.33	19.99
ATHLETE (ours)	$\mathbf{90.00}$ $(27/30)$	$\mathbf{86.67}$ $(26/30)$	$\mathbf{46.67}$ $(14/30)$	$\mathbf{53.33}$ $(16/30)$

TABLE III: Real-robot results under each method’s original evaluation protocol and metric definitions. AdaPT’s overall net-clearance rates are estimated by averaging the per-style products of hit and conditional net-clearance rates. Return denotes in-court return for LATENT and net clearance for AdaPT and ATHLETE. The motion-capture coverage areas are approximately

20\,\mathrm{m}\times 8\,\mathrm{m}

for AdaPT and

10.5\,\mathrm{m}\times 5.5\,\mathrm{m}

for ATHLETE. LATENT is evaluated on a full-size tennis court.

Property	Vid2Player3D	LATENT	AdaPT	ATHLETE (ours)
Existing public data reuse	$\checkmark$	—	$\checkmark$	$\checkmark$
No extra tennis MoCap demo collection	$\checkmark$	$\times$	$\checkmark$	$\checkmark$
Deployment with MoCap	—	$\checkmark$	$\checkmark$	$\checkmark$
No motion reference	$\times$	$\checkmark$	$\times$	$\checkmark$
No future-ball prediction	$\times$	$\checkmark$	$\times$	$\checkmark$
Single-level policy	$\times$	$\times$	$\times$	$\checkmark$

TABLE IV: Qualitative comparison of reported configurations. Dashes indicate unavailable or inapplicable entries.

Fig. 7: Real-world striking snapshots from the manual-throw experiment.

Explore similar work

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
TaskNPoint: How to Teach Your Humanoid to Hit a Backhand in Minutes
Jun 24, 2026Blake Werner, Ilona Demler, Pietro Perona +1RL for Robotics Robot Skill Learning
California Institute of Technology
Towards Professional Tennis Styles for Humanoid Robots with Adaptive Motion Planning and Tracking
Aug 20, 2026Tao Huang, Ruofei Liu, Xuchen Tang +15Robotics Robot Motion Planning
Noitom Robotics · Shanghai AI Laboratory · Shanghai Jiao Tong University +1

Mar 13, 2026J/K move · Enter open · S save

Learning Athletic Humanoid Tennis Skills from Imperfect Human Motion Data

Zhikai Zhang, Haofei Lu, Yunrui Lian, Ziqing Chen+11

Tsinghua University · Galbot Inc. · Shanghai AI Laboratory+2

Robotics Humanoid Robot Control Robot Imitation Learning

Open paper AskPDF arXiv

Human athletes demonstrate versatile and highly-dynamic tennis skills to successfully conduct competitive rallies with a high-speed tennis ball. However, reproducing such behaviors on humanoid robots is difficult, partially due to the lack of perfect humanoid action data or human kinematic motion data in tennis scenarios as reference. In this work, we propose LATENT, a system that Learns Athletic humanoid TEnnis skills from imperfect human motioN daTa. The imperfect human motion data consist only of motion fragments that capture the primitive skills used when playing tennis rather than precise and complete human-tennis motion sequences from real-world tennis matches, thereby significantly reducing the difficulty of data collection. Our key insight is that, despite being imperfect, such quasi-realistic data still provide priors about human primitive skills in tennis scenarios. With further correction and composition, we learn a humanoid policy that can consistently strike incoming balls under a wide range of conditions and return them to target locations, while preserving natural motion styles. We also propose a series of designs for robust sim-to-real transfer and deploy our policy on the Unitree G1 humanoid robot. Our method achieves surprising results in the real world and can stably sustain multi-shot rallies with human players. Project page: https://zzk273.github.io/LATENT/

Abstract

Explore similar work

Learning Athletic Humanoid Tennis Skills from Imperfect Human Motion Data

TaskNPoint: How to Teach Your Humanoid to Hit a Backhand in Minutes

Towards Professional Tennis Styles for Humanoid Robots with Adaptive Motion Planning and Tracking