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/
Figures & tables
Figure 1: Overview of LATENT . (a) We pre-train a motion tracker on collected imperfect human motion data. (b) We construct a correctable latent action space via online distillation. (c) We train a high-level policy to correct and compose latent actions for tennis task. (d) We transfer the policy to the real world via dynamics randomization and observation noise.
Figure 2: Visualization of generated ball trajectories.
Term
Weight
Term
Weight
Task
Approach to ball
10.0
Ball landing
30.0
Hit success
200.0
Regularization
High-level action
−5⋅10−3
Torque penalty
−2⋅10−5
Lower body action rate
−1.0
Whole body action rate
−0.5
Table 1: Reward terms used in LATENT. The reward terms are divided into three types: Task, Regularization, and Termination.
Figure 3: The motivation of Latent Action Barrier (LAB).
Term
Value
Robot
Feet friction
U(1.0,2.5)
Joint friction
×U(0.75,1.25)
Armature
×U(1.0,1.05)
Body mass
×U(0.75,1.25)
Body CoM
U(−0.05,0.05)
Table 2: Dynamics randomization used in LATENT. We randomize robot dynamics and tennis ball physics.
Method
Forehand
Backhand
Forecourt
Backcourt
SR↑
DE↓
Smth↓
Torque↓
SR↑
DE↓
Smth↓
Torque↓
SR↑
DE↓
Smth↓
Torque↓
SR↑
DE↓
Smth↓
Torque↓
PPO [ 26 ]
N/A
N/A
N/A
N/A
N/A
N/A
N/A
N/A
N/A
N/A
N/A
N/A
N/A
N/A
N/A
N/A
MotionVAE [ 44 ]
N/A
N/A
N/A
N/A
N/A
N/A
N/A
N/A
N/A
N/A
N/A
N/A
N/A
N/A
N/A
N/A
AMP [ 22 ]
41.32
4.71
46.52
13.79
36.81
5.22
37.68
15.74
49.48
4.20
36.16
18.26
41.73
4.85
41.77
19.20
ASE [ 21 ]
63.47
3.62
38.91
11.84
55.22
4.01
31.75
12.92
71.36
3.28
30.84
14.05
68.59
3.90
35.12
15.27
PULSE [ 19 ]
71.85
3.01
32.44
9.76
63.38
3.15
29.60
10.83
74.52
2.97
28.11
12.01
72.94
3.02
31.27
13.44
Table 3: Comparison with baseline methods. Bold numbers indicate the best performance.
Method
Forehand
Backhand
Forecourt
Backcourt
SR↑
DE↓
Smth↓
Torque↓
SR↑
DE↓
Smth↓
Torque↓
SR↑
DE↓
Smth↓
Torque↓
SR↑
DE↓
Smth↓
Torque↓
w/o Corr.
82.36
3.41
26.72
8.85
68.94
3.03
29.55
8.92
74.21
3.18
26.86
9.64
79.05
3.52
29.94
9.38
w/o LAB
93.12
2.17
37.64
12.53
76.96
2.21
31.77
12.11
84.73
1.96
30.59
10.94
87.44
2.05
32.08
13.62
Ours
96.52
1.32
25.61
7.40
82.10
1.89
27.76
7.71
86.35
1.73
25.51
7.47
89.80
1.55
27.20
7.60
Table 4: Ablation study. Bold numbers indicate the best performance.
Method
Forehand
Backhand
Forecourt
Backcourt
SR↑
DE↓
SR↑
DE↓
SR↑
DE↓
SR↑
DE↓
w/o DR
16.67%
5.60
25.00%
4.57
28.57%
3.59
14.29%
6.23
w/o ON
50.00%
3.28
0.00%
N/A
28.57%
3.69
37.50%
3.97
Ours
90.90%
3.15
77.78%
3.89
88.89%
3.57
81.82%
3.80
Table 5: Real-world performance. Bold numbers indicate the best, and underlined numbers indicate the second.
Appendix figures & tables4 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 4: Robot movement coverage during consecutive ball returns. Heatmaps of the robot’s global positions accumulated over different numbers of consecutive ball returns (8, 16, 80, and 400). The learned policy enables effective court coverage and adaptive repositioning during consecutive rallies.
Figure 5: Robot-robot self-play evaluation in simulation. (a) Visualization of two robots performing self-play on opposite sides of the court. (b) Distribution of number of consecutive rallies over 50 random games.
Figure 6: Different hitting events in simulation. The red curves denote incoming ball trajectories, the robot posture corresponds to the moment of ball contact, and the green curves indicate the trajectories of the returned balls.
Figure 7: Close-up real-world examples. Representative frames from real-world rallies showing diverse tennis return behaviors, including forehand and backhand strokes, dynamic footwork, and coordinated whole-body motion.