Organizations: University of Science and Technology of China · State Key Laboratory of General Artificial Intelligence, BIGAI · Shanghai Jiao Tong University · Peking University
Humanoid soccer requires a robot to seamlessly combine agile locomotion, ball manipulation, and impact-rich contact dynamics across a continuous, dynamic behavioral spectrum. Yet existing approaches either cover only a limited set of skills, or forego motion priors entirely and produce unnatural behaviors, leaving the combination of a unified skill set and human-like motion quality unaddressed. We present soccerbot, a hierarchical reinforcement learning framework that resolves both limitations through a structured latent motion prior distilled from large-scale human soccer motion capture data. By constraining policy learning within this dynamically plausible motion manifold, a single skill-conditioned policy naturally learns to perform moving, dribbling, trapping, and kicking in a unified formulation, while inherently producing human-like motion and fluid transitions during commanded skill switching. Augmented by robustness-oriented simulation design, soccerbot transfers zero-shot to the Unitree G1 humanoid robot, demonstrating agile, natural, and robust execution of the full soccer skill repertoire under real-world conditions.
Figures & tables
Figure 2: Overview of SoccerBot . The framework consists of three stages: (1) a motion tracking expert is trained on retargeted human soccer motion data; (2) the tracking expert is distilled into a structured latent motion prior via a variational encoder-decoder framework; (3) a unified skill-conditioned policy learns to perform moving, dribbling, trapping, and kicking via residual latent control over the frozen prior, with transition state initialization enabling robust cross-skill composition. Real-world deployment uses an external motion capture system for ball state estimation.
Method
Kicking
Dribbling
Trapping
Moving
Motion alignment
SR ↑
VE ↓
DE ↓
SR ↑
VE ↓
DE ↓
ResT ↓
SR ↑
StopT ↓
DistErr ↓
SR ↑
VE ↓
DE ↓
Train ref. ↓
Post-train ref. ↓
PPO
93.96
8.68
4.35
0.00
99.81
163.83
5.00
94.64
0.87
0.31
99.96
6.07
1.77
0.328
0.575
AMP [ 13 ]
87.48
10.32
7.26
46.36
37.05
47.88
4.09
23.08
5.49
3.07
99.96
9.43
4.21
0.170
0.308
ASE [ 12 ]
0.00
99.52
169.79
0.00
99.67
163.64
5.00
49.24
3.46
1.58
58.60
28.59
26.62
0.292
0.409
LATENT [ 28 ]
91.00
8.50
6.80
85.24
12.69
7.21
1.92
85.92
1.53
0.64
100.00
9.67
2.71
0.231
0.603
Ours
97.04
7.26
4.76
89.88
8.79
4.24
1.36
83.08
1.43
0.49
100.00
4.95
5.38
0.226
0.554
Table 1: Main simulation results on four humanoid soccer skills.
Figure 3: Latent skill-space construction under a joint t-SNE projection. Samples from zprior , ztask , and the final latent z are embedded together and shown in separate panels. Colors indicate skill labels. The prior latent captures shared state-conditioned motor modes, while the task residual and final latent reveal increasingly skill-dependent structure.
Method
Kicking
Dribbling
Trapping
Moving
VelErr (%)
DirErr ( ∘ )
VelErr (%)
DirErr ( ∘ )
ResT (s)
StopT (s)
DistErr (m)
VelErr (%)
DirErr ( ∘ )
Ours (sim)
8.12
4.09
8.55
3.77
1.21
0.78
0.31
19.14
5.85
Ours (real)
47.38
10.41
30.17
16.49
2.37
1.84
1.07
21.53
13.74
Table 4: Simulation and real-robot results on four humanoid soccer skills.
Appendix figures & tables3 assets
Supplementary material from the paper’s appendix.
Appendix
Group
Term
Dim.
Corruption
Policy/Critic
Command
3
None
Policy/Critic
Ball position
3
[−0.02,0.02] m
Policy/Critic
Ball velocity
3
[−0.1,0.1] m/s
Policy/Critic
Skill one-hot
4
None
Policy/Critic
Base angular velocity
3
[−0.05,0.05] rad/s
Critic only
Base linear velocity
3
None
Appendix
Table A.1: Observation layout for the unified soccer skill policy and critic. The policy observation concatenates all listed policy terms and applies corruption where specified. The critic uses the corresponding uncorrupted terms, additionally includes base linear velocity, and disables corruption. For the Unitree G1, there are 29 actuated joints, yielding 106 policy dimensions and 109 critic dimensions.
Table A.2: Reward terms for the unified skill environment. Idrib , Ikick , Imove , and Itrap denote active-skill indicators; vc , vb , and vr denote the command, ball, and root velocities, respectively; pbr denotes the ball position in the robot frame; df denotes the nearest foot–ball distance; and g(⋅) denotes implementation-specific gates, including contact phase, possession, standing-command mode, and reward ramping.
Term
Unit
Range
Robot static friction
–
[0.3,1.6]
Robot dynamic friction
–
[0.3,1.2]
Robot restitution
–
[0.0,0.15]
Soccer static friction
–
[0.35,0.75]
Soccer dynamic friction
–
[0.25,0.55]
Soccer restitution
–
[0.45,0.85]
Appendix
Table A.3: Domain randomization and reset perturbations. Ranges are sampled uniformly unless otherwise stated.