Humanoid kicking requires coordinated whole-body motion and precise contact, while a banana kick demands contact mechanics that generate ball spin and aerodynamic curvature. Motion imitation provides a reliable ordinary-kick prior, but reinforcement learning may improve shot speed and placement accuracy without changing the underlying kicking technique. Adapting this prior to a qualitatively different contact-rich skill can fail even when the reward is dense and optimization remains stable. The failure occurs when the task objective is locally flat over the current policy's responses. We term this condition first-order learning starvation. To address it, we propose response-informed skill evolution (RISE), a closed-loop objective-continuation method for policy adaptation. RISE ranks bounded objective changes using response sensitivity estimated from cached rollouts and accepts updates only when they produce verified response progress while preserving kicking reliability. Our analysis shows that rescaling a saturated spin reward cannot recover first-order sensitivity at zero spin, whereas adapting coupled contact responses can provide a learnable path to spin generation. We integrate RISE into a humanoid kicking pipeline under calibrated contact and Magnus-force aerodynamics, and sim-to-real transfer. Experiments show that RISE evolves the ordinary kick into a high-spin curved kick with 11.55 rad/s mean ball spin, improves the mean evaluation score by 19.8% over a learning-progress curriculum, and raises joint target attainment from 15.2% to 50.9%. Ablations and response diagnostics support the mechanism, while 30 motion-capture-recorded physical trials demonstrate consistent hardware transfer of the learned curved kick. Project website: https://haozhang-thu.github.io/bananakick/
Figures & tables
Fig. 2: From motion prior to new physical skill: imitation provides a reliable ordinary kick while RISE evolves its contact regime under calibrated contact dynamics and Magnus-force aerodynamics. Cached response sensitivities guide bounded objective proposals, and response progress with correct-foot reliability determines whether each policy update is retained before direct hardware transfer.
Item
Value
Item
Value
Robot
G1
Seeds
5
Prior
Ordinary
Bank
4096
Offset
±5 cm
Bootstrap
10 k
Actor
Mean
Reliability
Correct-foot
δ0
.35
δ bounds
[.02,.60]
γ↑/γ↓
1.5/.5
β/κ/qmin
1/.15/.85
TABLE I: Experimental settings.
Fig. 3: Simulation and sim-to-real comparison: (a,b) show the simulated kick and straight-versus-banana trajectories, while (c,d) compare physical and simulated contact-side sweep and follow-through; quantitative flight trajectories are reported in Fig. 11 .
Fig. 4: Cross-method evaluation over five seeds and 4,096 matched contexts per seed: (a) fixed-objective score and (b) normalized response regret. LP-Curr. denotes the learning-progress curriculum; No-Rank and No-FB remove sensitivity ranking and closed-loop feedback, and Direct denotes direct-final PPO.
Fig. 5: Paired comparisons quantify the RISE effect through (a) score differences and (b) patch-precision improvement versus spin change; abbreviations follow Fig. 4 .
Fig. 6: Straight- and curved-kick executions show approach, contact, and follow-through, with overlaid paths illustrating their post-contact differences; calibrated flight measurements are reported in Fig. 11 .
Fig. 7: Physical-response evolution across accepted RISE checkpoints: (a) prior-to-final gain over ball offsets, (b) normalized response coordinates, (c) consecutive paired-score increments, and (d) patch-error–spin distributions relative to the joint thresholds.
Method
Score
Gap ↓
Coverage (%)
CVaR10
Regret ↓
RISE win (%)
RISE
1093.7 [1092.1,1095.3]
431.3
50.88
975.5
0.0149
—
LP curriculum
912.8 [909.3,916.3]
612.2
15.23
710.9
0.1427
94.21
RISE w/o ranking
729.6 [724.2,735.0]
795.4
0.073
414.2
0.0846
99.34
RISE w/o feedback
163.2 [162.1,164.3]
1361.8
0.000
86.4
1.2135
100.00
Direct PPO
71.7 [71.5,72.0]
1453.3
0.000
58.1
1.9466
100.00
TABLE II: Cross-method comparison over five independent seeds under the fixed evaluation score; brackets denote 95% matched-context bootstrap intervals.
Method
Spin
Patch error ↓
Tangential speed
Ball speed
Alignment
RISE
11.549
2.274
8.059
8.711
0.899
LP curriculum
11.553
3.575
7.644
8.619
0.919
RISE w/o ranking
13.210
4.631
6.584
8.418
0.912
RISE w/o feedback
1.887
11.099
6.172
10.073
0.764
Direct PPO
0.000
15.776
4.988
8.296
0.845
TABLE III: Simulated terminal-response diagnostics for the five policies. Spin is in rad/s, patch error in cm, and speeds in m/s.
Fig. 8: Sensitivity-scale diagnostics across accepted RISE checkpoints: (a) patch excitation over admissible scales, (b) incumbent versus response-matched reference scales, (c) scale mismatch versus subsequent channel change, and (d) excitation and deformation for accepted and rejected candidates.
Fig. 9: Physical curved-kick execution through approach, contact, and follow-through, with overlays illustrating the observed post-contact ball path; quantitative trajectories are reported in Fig. 11 .
Fig. 10: Measured response-space starvation and repair: (a) reachable patch-error support versus reward sensitivity under admissible scales and (b) response distributions before and after correction.
Fig. 11: Frozen-policy sim-to-real validation: (a) simulated and motion-capture-recorded trajectories; and (b) target-plane errors. Here, 20 simulated and 20 physical trajectories, randomly sampled from the original dataset, are shown for clarity.
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
Elite humanoid soccer shooting requires whole-body stability, high-impulse whole-body interactions, and accuracy to targets. Motion tracking-driven reinforcement learning (RL) provides stability in whole-body movement coordination, but a fixed reference makes it hard to adapt to varied ball positions and strike timings; in contrast, task reward-driven RL struggles to explore and discover valid kicks from scratch. We therefore introduce RoboNaldo, a three-stage motion-guided curriculum RL framework for high-impulse humanoid interaction. A single human-kick reference is used as a scaffold and progressively shifts optimization towards shooting performance. The curriculum first learns a stable whole-body kicking prior, then adapts the kick to free-kick settings where the ball is stationary at random positions, and finally extends it to moving-ball shooting through a locomotion-command and kick-trigger interface. A high-level heuristic planner controls this interface during training, while alternative high-level controllers can drive the same low-level policy at inference. In simulation, RoboNaldo demonstrates free-kick shot error 48.6% lower and shoot velocity 2.96x than prior work baselines. In real world on a Unitree G1 with onboard perception, RoboNaldo attains 0.73 m and 0.86 m average target shooting error from 3 m away in free-kick and moving-ball cases, accordingly. And the post-contact ball velocity reaches 13.10 m/s, which is 59-71% of reported professional open-play shot speed. Project page: https://opendrivelab.com/RoboNaldo.
Yichao Zhong, Yidan Lu, Yuhang Lu +9
The University of Hong Kong · Archon Robotics · The Chinese University of Hong Kong
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.