Humanoid soccer contact skills require more than producing high-impact foot-ball contacts: the robot must close the loop over perception, approach, alignment, impact, and recovery while its own motion induces substantial viewpoint changes, frequent loss of the ball from view, and uncertain contact outcomes. In this work, we ask a compact yet stricter question: can a humanoid learn soccer contact skills using only a head-mounted depth image, proprioceptive history, and an optional low-dimensional task command, and directly output 25-DoF joint PD targets without extra runtime perception or planning modules? To this end, we propose DAVIS, a depth-only end-to-end framework for humanoid soccer skills that learns visibility-aware auxiliary geometry during training, and combines GT-to-prediction annealing, task curricula, and AMP-style motion priors to smoothly bridge privileged supervision and real deployment. Built on this framework, we instantiate representative soccer contact skills, including goal-directed shooting and directional dribbling, through task-specific definitions of objects, commands, rewards, and curricula, and validate them through simulation, Noetix E1 real-robot experiments, and ablations.
Figures & tables
Fig. 2: Overview of our DAVIS framework. The observation stream contains the command ct , a single 168×80 depth image , proprioceptive history, and privileged simulation states. The structured-depth module encodes the depth input into a 32-D latent, predicts ball/goal geometry and visibility confidence through auxiliary heads , forms confidence-gated geometry features, and combines them with a 64-D HIM/history feature. The actor receives the command, depth latent, gated geometry/confidence features, proprioceptive history, and HIM feature, and outputs a 25-D position residual for the PD controller. The auxiliary losses, GT-to-prediction annealing, PPO update, task reward, and style reward use simulation-only signals from IsaacLab rollouts, including ball, goal, and visibility ground truth. Dashed boxes in the structured-depth module indicate optional task-specific modules.
Fig. 5: Qualitative examples of learned soccer skills. In simulation: (a) curved and penalty kicks, (b) cone-avoidance dribbling, (d) dribbling across a soccer field, and (f) stopping a rolling ball, where red and blue arrows indicate robot and ball motion, respectively, and the purple box marks the stopped ball. On the real robot: (c) foot–ball contact during dribbling, with arrows indicating ball motion, and (e) a close-up dribbling sequence.
Goal-Directed Shooting
Directional Dribbling
Setting
Metric
Easy
Med.
Hard
Total
Setting
Metric
N=3
N=5
N=7
60∘
90∘
120∘
60∘
90∘
120∘
60∘
90∘
120∘
Sim.
Penalty SR ↑
0.95
0.96
0.56
0.85
Sim.
SR ↑
0.71
0.77
0.65
0.66
0.72
0.55
0.68
0.64
0.43
Free-kick SR ↑
1.00
0.88
0.71
0.85
Evel↓
0.69
0.74
0.91
0.74
0.73
0.80
0.71
0.81
0.97
–
–
–
–
–
Rbib↑
0.73
0.74
0.64
0.76
0.75
0.70
0.76
0.72
0.60
Real
Penalty SR ↑
0.68(26/38)
0.57(28/49)
0.55(33/60)
–
Real
SR ↑
Easy: 0.79(22/28)
Medium: 0.68(19/28)
Hard: 0.61(17/28)
TABLE I: Quantitative results for goal-directed shooting and directional dribbling in simulation and on the real robot. Shooting entries and real-robot dribbling entries report success rate (SR) on a 0–1 scale; simulation dribbling reports SR, velocity error ( Evel ), and ball-in-band ratio ( Rbib ) over 900 repeated-S slalom trials. Arrows indicate the preferred direction, and bold marks the best result within each fixed- N simulation-dribbling group. For shooting, a success requires the ball to cross the goal line in one kick. For dribbling, a success requires the robot to remain upright throughout execution and both the robot and ball to finish within 1 m of the target.
Fig. 6: (a) Free-kick and Penalty-kick performance in simulation. Each region uses over 50 trials. Darker colors indicate higher success rates, and dashed lines mark difficulty regions. (b) Footstep trajectories during shooting. The policy shows a smooth gait and consistent speed pattern.
Fig. 7: Performance of repeated S-slalom in simulation. (a) Statistical results of 100 S-slalom trajectories. (b) Foot-contact record from a representative successful trajectory.
Method
Penalty kick
Free kick
Repeated-S slalom
SR ↑
Eang↓
rcontact↑
SR ↑
Eang↓
rcontact↑
SR ↑
Evel↓
rbib↑
DAVIS
0.84
23.29
99.75
0.80
30.18
99.27
0.70
0.79
0.74
(a) Ablation
DAVIS-w/o-curriculum
0.62
27.40
32.53
0.55
34.42
87.67
0.51
0.77
0.69
DAVIS-auxiliary head-w/o-anneal
0.32
54.79
84.87
0.22
76.20
87.92
0.31
0.83
0.66
DAVIS-auxiliary head-w/o-geometry
0.20
53.53
87.33
0.15
69.47
78.13
0.34
0.91
0.54
TABLE II: Simulation ablations and comparison. (a) Ablations. Penalty-kick and free-kick results use over 900 shooting trajectories, and repeated-S slalom uses over 100 dribbling trajectories. Metrics are success rate (SR), body-to-goal angular error ( Eang ), successful ball-contact ratio ( rcontact ), velocity error ( Evel ), and time in the control band ( rbib ). The full method provides the best overall balance across tasks and metrics. (b) Comparison between our end-to-end learned head control (Active vision) and a heuristic head controller (Heuristic) that gazes at the predicted ball. The heuristic performs comparably to active vision in static-object tasks (e.g., penalty kicks), while active vision shows a clear advantage in highly dynamic dribbling scenarios.
Appendix figures & tables19 assets
Supplementary material from the paper’s appendix.
Appendix
Metric
Definition
Success rate (SR)
The main paper defines task-level success: shooting requires the ball to cross the goal line in one kick, while dribbling requires the robot to remain upright and both the robot and ball to finish within 1 m of the target. This appendix expands that definition for evaluation protocols: simulated dribbling additionally requires visiting all waypoints in order (Sec. IV-A ), whereas the real-world protocol applies the same endpoint and uprightness criteria to each trial.
Valid contact ratio ( rcontact )
Fraction of shooting trials with a valid foot–ball contact and goal-directed ball exit.
Angular error ( Eang )
Heading error between the robot body and the robot-to-goal direction.
Velocity error ( Evel )
Mean error between commanded and actual ball velocity during dribbling.
Ball-in-band ratio ( Rbib )
Fraction of engaged dribbling time in which the ball stays inside the controllable band.
Turn bias ( Bturn )
∣θˉachieved−θcmd∣/θcmd over scored rounds under the Dribble Master turning protocol; the statistic behind its reported turning error.
Appendix
TABLE III: Metric definitions used throughout the evaluation.
Fig. 8: Multi-position Shooting Simulation and Real Machine Difficulty Level Classification
Field
Definition
Value
Task grid
Repeated-S slalom with trajectory turns N and turn angle α
N∈3,5,7 , α∈60,90,120∘
Waypoints
N+2 turning waypoints plus N+1 midline gate points
2N+3 total
Geometry
Fixed arm length for waypoint generation
L=4.0 m
Command
Waypoint attraction and pole repulsion, snapped to 12 direction bins
Fig. 9: Trajectory visualization for the repeated-S slalom dribbling benchmark. Each subplot corresponds to one task configuration (N,α) , where N is the number of turns and α is the turning angle. The colored curve shows the mean successful ball trajectory with speed-based coloring, and the shaded region indicates the 15–85% success band. The start point and target region are marked for reference. The title of each subplot reports the success rate and the number of successful rounds.
Field
Dribble Master [ 27 ]
DAVIS reproduction
Robot / policy input
Booster T1; detector ball position + in-view flag
Noetix E1; depth image
Command
global ball-velocity vector, issued by a human operator
direction only, scripted (potential field for reaching, fixed heading for turning)
Reaching / obstacle
unreported distance; success = final error <1 m
goal at d∈{3,5,8} m; success = minimum ball–goal distance <τ , τ∈{0.5,1.0,1.5} m; round ends at 0.5 m or after max(20,6d) s
Obstacle
physical; not perceived by the policy
virtual pole at the path midpoint; not perceived by the policy
Turning
±45∘/90∘ on entering a 0.4 m circle at (1.5,0) m
±30/45/60/90∘ on 1.1 m of forward progress (where a straight path enters that circle)
Trials
5 rollouts per turn condition (sim); 15 per task (real)
50 rounds per condition (sim), Wilson 95% CI
Appendix
TABLE V: Dribble Master’s published dribbling protocol and our reproduction of it on DAVIS.
Task
d (m)
RR0.5
RR1.0
RR1.5
time (s)
Target reaching
3
0.88 [0.76, 0.94]
0.88 [0.76, 0.94]
0.90 [0.79, 0.96]
3.0±0.6 (44)
5
0.82 [0.69, 0.90]
0.84 [0.71, 0.92]
0.84 [0.71, 0.92]
5.7±0.5 (42)
8
0.78 [0.65, 0.87]
0.80 [0.67, 0.89]
0.80 [0.67, 0.89]
9.4±1.1 (40)
Single obstacle
3
0.92 [0.81, 0.97]
0.92 [0.81, 0.97]
0.92 [0.81, 0.97]
4.8±1.1 (46)
5
0.94 [0.84, 0.98]
0.96 [0.87, 0.99]
0.96 [0.87, 0.99]
6.4±1.1 (48)
8
0.94 [0.84, 0.98]
0.94 [0.84, 0.98]
0.94 [0.84, 0.98]
9.9±1.1 (47)
Appendix
TABLE VI: Target-reaching and single-obstacle results under the Dribble Master protocol (DAVIS simulation, n=50 rounds per row, Wilson 95% CI). Success is reported at three tolerances, together with the time from first ball contact to the 1 m tolerance (mean ± SD over the rounds that got there, count in parentheses); a round ends when the ball first comes within 0.5 m. Dribble Master’s reaching and obstacle values come from real-robot trials with n=15 at an unreported distance.
Turn command
DAVIS
Dribble Master
scored
θˉ±σ ( ∘ )
bias (%)
per-round (%)
θˉ ( ∘ )
bias (%)
30 ∘ L
40
+28.6±20.5
4.8
52.9
–
–
30 ∘ R
39
−25.5±15.3
15.1
41.5
–
–
45 ∘ L
44
+39.3±14.9
12.6
27.6
+43.58
3.16
45 ∘ R
44
−44.8±18.2
0.5
32.1
-46.44
3.20
60 ∘ L
43
+60.4±19.5
0.6
26.3
–
–
Appendix
TABLE VII: Turning results under the Dribble Master protocol (DAVIS simulation, n=50 rounds per condition). Achieved heading change is reported as mean ± SD over scored rounds, together with the bias of the mean. Dribble Master’s turning values come from MuJoCo with 5 rollouts per condition.
Fig. 10: Ablation on active vision.
Fig. 11: Multi-directional shooting real-world example
Difficulty
Distance (m)
Target Angles
Easy
2.5
−30∘ , −5∘ , +5∘ , +30∘
Medium
3.0
−30∘ , −5∘ , +5∘ , +30∘
Hard
2.5–3.0
−45∘ , +45∘
Appendix
TABLE VIII: Target point configurations and difficulty levels used in the real-world shooting evaluation.
Fig. 12: Foot snapshot in dribbling experiment.
Item
Configuration
Actor sensory input
Single-frame depth image + 5-step proprioceptive history
Optional command
Low-dimensional task command if enabled
Shooting proprioception
81-D
Dribbling proprioception
84-D
Depth encoder
LightDepthEncoder
Depth latent
32-D
Appendix
TABLE IX: Observation and network configuration.
Item
Value
Rollout length
24 steps per environment
PPO epochs
5
Mini-batches
4
Learning rate
1×10−3 , adaptive schedule
Discount factor
γ=0.99
GAE parameter
λ=0.95
Appendix
TABLE X: PPO and AMP-HIM optimization settings.
Fig. 13: Loss curve of GT-annealing and w/o GT-annealing
Reward Terms
Components
Weight
Approach and chase
Approach ball reward
3.0
Adaptive velocity reward
2.0
Active vision
Head track ball reward
3.0
Keep in horizon reward
-0.5
Body follow head reward
2.0
Ball–goal alignment
Robot behind ball reward
1.0
Appendix
TABLE XI: Shooting reward summary.
Group
Terms
Weight / setting
Commanded progress & heading
Ball command reward
5.0
Body command reward
0.4
Ball control
Dense re-approach reward
2.0
Out-of-band penalty
-1.0
Active vision
Ball in view reward
1.0
Body follow head reward
0.6
Appendix
TABLE XII: Dribbling reward and regularization summary.
Perturbation
Range / value
Valid depth range
0.3–5.0 m
Distance-proportional depth noise
0.01
Gaussian depth-noise coefficient
0.005
Salt-and-pepper probability
0.01
Stripe dropout probability
0.005
Edge-drag probability
0.01
Appendix
TABLE XIII: Representative training randomization ranges for sim-to-real transfer.
Fig. 14: Power shoot real-world example
Fig. 15: Auxiliary loss for the ball and obstacles. The prediction heads for both objects converge during training.
Recent advances in humanoid robotics have highlighted the importance of deployable loco-manipulation skills. Dribbling a soccer ball while evading active opponents requires simultaneous balance, precise ball control, and awareness of a dynamic adversary under onboard sensing and real-time constraints. Existing approaches typically separate perception and motion, which can be effective in controlled settings but may fail under occlusions, fast ball movements, and complex opponent interactions, since perception is not directly optimized for control. We propose an integrated approach in which a temporal depth encoder is embedded into a reinforcement learning policy through a task-specific projection layer. We apply this framework to a simulated Booster T1 humanoid robot and show that it is possible to learn vision-based, opponent-aware dribbling directly from depth observations, without explicit state estimation or privileged scene information. The learned policy achieves 100% success in nominal target-driven dribbling and 96% success with a single static obstacle, while reaching 46% success against an actively moving ball-attacker opponent. These results demonstrate that the proposed framework supports robust vision-based dribbling in nominal and moderately dynamic settings, and provides a strong foundation for handling more challenging moving-adversary scenarios.
Dept. of Computer, Control, and Management Engineering Sapienza University of Rome, Rome (Italy) · Institut de Robòtica i Informàtica Industrial (CSIC-UPC), C/ Llorens i Artigas 4-6, 08028, Barcelona, Spain · École Polytechnique Fédérale de Lausanne (EPFL), Lausanne, Switzerland
Humanoid soccer is a challenging testbed for dynamic whole-body control, requiring robots to coordinate balance, locomotion, object interaction, and skill switching over long horizons. Existing humanoid sports methods often rely on task-specific multi-stage pipelines, making it difficult to jointly learn and compose multiple object-interactive skills within a single deployable policy. To address this, we present SkillX, a unified reinforcement learning framework that learns and composes multiple atomic soccer skills through a single command-conditioned policy. SkillX integrates three core designs: skill-specific adversarial motion priors, skill-specific critics, and an object-aware temporal encoder, enabling the robot to execute atomic skills and transition among them such as dribbling, trapping, and shooting. Experiments in simulation and on a real Noetix E1 humanoid demonstrate robust multi-skill execution, long-horizon skill composition, and successful sim-to-real deployment.
Recent humanoid soccer systems make motion tracking the substrate and derive locomotion from it, typically by steering a motion-reference anchor toward the ball. This yields strong shooting results, but locomotion is trained only on the narrow, deterministic command distribution ball approach induces, never evaluated as a capability in its own right. We invert the stack: a general, command-conditioned locomotion policy is trained first as the substrate, and N motion-guided kicking skills are added on top as task-gated layers, so the reachable gait space is set by the locomotion curriculum rather than any reference clip. Because every skill starts from and returns to this same commandable state, locomotion also becomes a composition hub (O(N) transitions rather than O(N^2)), and post-strike stabilisation is handed back to the trained controller rather than scripted per clip. We instantiate this on a 29-DoF Unitree G1 with seven retargeted kicking skills spanning 259.5 degrees of nominal aim direction, including lateral, rearward and weak-foot strikes a single forward-facing reference cannot express, and report shooting accuracy alongside command-tracking, terrain and push-recovery results with the full skill library attached, an axis prior humanoid soccer systems do not report. The library is validated on hardware across forward, lateral, rearward and commanded approaches.
Abu Hanif Muhammad Syarubany, Jaehyun Jang, Hwanhee Kim +3
School of Electrical Engineering, KAIST, Daejeon, Republic of Korea