Locomotion-Grounded Humanoid Soccer: Task-Gated Reinforcement Learning of a Multi-Directional Kicking Library

Abu Hanif Muhammad Syarubany; Jaehyun Jang; Hwanhee Kim; Kyuwon Kim; Seungyeon Ryu; Chang D. Yoo

cs.ROSep 30, 2026

Locomotion-Grounded Humanoid Soccer: Task-Gated Reinforcement Learning of a Multi-Directional Kicking Library

Authors: Abu Hanif Muhammad Syarubany, Jaehyun Jang, Hwanhee Kim, Kyuwon Kim, Seungyeon Ryu, Chang D. Yoo

Organizations: School of Electrical Engineering, KAIST, Daejeon, Republic of Korea

Abstract

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.

Figures & tables

Fig. 1 : Substrate inversion. Prior work derives locomotion from a ball-steered anchor; ours is the commandable base.

Fig. 2 : Training pipeline and deployed architecture. Training (left). Stages A–C are Stages 1–3 of Sec. IV : locomotion first and alone, then motion tracking, then motion-guided kicking, with specialist distillation (Stage 4) below. Deployment (right). The same distilled, task-mode-gated network, with the perception hardware of Sec. V supplying ball position.

		Locomotion substrate		Skill repertoire				Handoff
Method	Platform	Free	Omni	Multi	Motion	Lateral	Rearward	Loco	Kick
		command	directional	skill	guided	target	target	$\to$ Kick	$\to$ Loco
Haarnoja et al. [ 1 ]	Miniature	✗	$\circ$	✗	✗	✗	✗	✓	✓
STOFT [ 19 ]	Bipedal	✗	✗	✓	✗	✓	✗	✓	✓
Reactive [ 6 ]	Humanoid	✗	✗	$\circ$	✓	$\circ$	$\circ$	✓	$\circ$
Striker [ 4 ]	Humanoid	$\circ$	✗	✗	✗	✗	✗	✓	✓

TABLE I: Capability comparison against prior humanoid soccer systems. Lateral and rearward targets require the strike bearing to be an explicit command input spanning that range; systems that reach such bearings only emergently score

\circ

Locomotion group ( $o_{loco}$ )		Kicking group ( $o_{kick}$ )
Term	Dim	Term	Dim
Base angular velocity	$3$	Motion reference $(q,\dot{q})$	$58$
Projected gravity	$3$	Reference orientation	$6$
Command $(v_{x},v_{y},\omega_{z})$	$3$	Base angular velocity	$3$
Joint position	$29$	Joint position	$29$
Joint velocity	$29$	Joint velocity	$29$

TABLE II : Actor observation space,

261

dimensions.

Term	$w$	Term	$w$
Motion tracking
Global root pos.	$1.0$	Global root orient.	$0.5$
Relative body pos.	$1.0$	Relative body orient.	$1.0$
Global body lin. vel.	$1.0$	Global body ang. vel.	$1.0$
Global feet lin. vel.	$1.0$	Strike joint null-sp.	$1.0$
Strike leg null-sp.	$1.0$

TABLE III : Kick-mode reward terms and weights. Twenty-two standard safety and regularisation terms (weights

-100

50

) are omitted for space.

Skill	Foot	Nominal	Reachable band
		bearing	( $\theta_{\max}=\pm 15^{\circ}$ )
skill_1	right	$-1.32^{\circ}$	$[-16.3,+13.7]$
skill_2	right	$+137.80^{\circ}$	$[+122.8,+152.8]$
skill_3	left	$-121.74^{\circ}$	$[-136.7,-106.7]$
skill_4	right	$-56.94^{\circ}$	$[-71.9,-41.9]$
skill_5	right	$+48.39^{\circ}$	$[+33.4,+63.4]$

TABLE IV : Skill library geometry.

Method	Bearing range	Dist. (m), tol.	suc.@ $0.5$ m (%)	suc.@ $1.0$ m (%)	Shot err. (m)	$v^{\max}_{\mathrm{ball}}$ (m/s)	Hit (%)	Alive (%)
PPO [ 24 ]	forward	n/r	$0.0{\pm}0.0$	$0.0{\pm}0.0$	$4.721{\pm}.007$	$0.780{\pm}.003$	$0.0{\pm}0.0$	$0.0{\pm}0.0$
AMP [ 11 ]	forward	n/r	$0.9{\pm}0.1$	$3.3{\pm}0.1$	$3.733{\pm}.004$	$1.771{\pm}.051$	$38.6{\pm}2.8$	$41.6{\pm}2.3$
PAiD [ 3 ]	forward	n/r	$8.2{\pm}0.7$	$19.5{\pm}0.5$	$1.850{\pm}.026$	$4.986{\pm}.017$	$67.7{\pm}0.3$	$67.3{\pm}0.5$
RoboNaldo [ 2 ] , stage 2, free-kick	forward, $\pm 38.7^{\circ}$	$5.0$ , n/r	$28.8{\pm}1.3$	$65.5{\pm}0.6$	$0.899{\pm}.012$	$\mathbf{14.792{\pm}.011}$	$82.1{\pm}0.3$	$\mathbf{100.0{\pm}0.0}$
RoboNaldo [ 2 ] , stage 3, free-kick	forward, $\pm 38.7^{\circ}$	$5.0$ , n/r	$36.3{\pm}0.1$	$66.5{\pm}0.2$	$1.055{\pm}.012$	$14.178{\pm}.034$	$81.1{\pm}0.1$	$99.7{\pm}0.1$
RoboNaldo [ 2 ] , stage 3, moving-ball	forward, $\pm 38.7^{\circ}$	$5.0$ , n/r	$32.4{\pm}0.6$	$63.3{\pm}0.6$	$1.131{\pm}.004$	$13.875{\pm}.004$	$79.2{\pm}0.1$	$98.8{\pm}0.1$

TABLE V : Shooting quality in context. PPO, AMP and PAiD rows are the free-kick simulation baselines as reported in [ 2 ] .

Fig. 3 : Hardware experiments on the Unitree G1. Top: commanded locomotion and two library strikes, each a multi-exposure composite of one trial. Bottom: a kick-to-locomotion handoff, from the mode switch through contact to locomotion resuming with no scripted stabilisation tail (Sec. VI-C ).

Skill	$1$	$2$	$3$	$4$	$5$	$6$	$7$	mean
Suc.@ $0.5$ (%)	$10.2$	$0.8$	$20.6$	$25.8$	$36.6$	$7.6$	$6.4$	$15.4$
Suc.@ $1.0$ (%)	$22.0$	$3.2$	$40.4$	$51.0$	$67.4$	$23.8$	$17.8$	$32.2$
Shot err. (m)	$1.90$	$2.92$	$0.98$	$1.10$	$0.81$	$1.89$	$2.15$	$1.68$
$v^{\max}_{\mathrm{ball}}$ (m/s)	$3.63$	$2.22$	$2.84$	$2.82$	$2.04$	$2.36$	$4.13$	$2.86$
Hit (%)	$100.0$	$92.0$	$85.2$	$100.0$	$100.0$	$100.0$	$100.0$	$96.7$
Alive (%)	$99.6$	$100.0$	$98.6$	$100.0$	$100.0$	$100.0$	$97.4$	$99.4$

TABLE VI : Per-skill performance of the unified policy.

Panel of Figure 10 — Fig. 4 : What the library covers, and how far the ball may sit from where a skill expects it.

Skill	Teacher	$N{=}2$	$N{=}4$	$N{=}6$	$N{=}7$
skill_1	$44.3$	$37.8\,(-6.5)$	$22.5\,(-21.8)$	$29.0\,(-15.3)$	$10.2\,(-34.1)$
skill_2	$19.6$	$14.2\,(-5.4)$	$8.0\,(-11.6)$	$5.3\,(-14.3)$	$0.8\,(-18.8)$
skill_3	$31.8$	–	$31.8\,(+0.0)$	$19.2\,(-12.6)$	$20.6\,(-11.2)$
skill_4	$30.0$	–	$20.7\,(-9.3)$	$8.8\,(-21.2)$	$25.8\,(-4.2)$
skill_5	$46.4$	–	–	$30.2\,(-16.2)$	$36.6\,(-9.8)$
skill_6	$41.8$	–	–	$15.3\,(-26.5)$	$7.6\,(-34.2)$

TABLE VII : Negative transfer: specialist vs. unified.

Direction	Trigger condition	Alive	Hit	Suc.@ $0.5$	Suc.@ $1.0$
Kick $\to$ Loco	end of authored clip	$99.1$	n/a	n/a	n/a
Kick $\to$ Loco	random tick, incl. strike	$98.8$	n/a	n/a	n/a
Loco $\to$ Kick	from $v\!\approx\!0$ state	$99.3$	$95.3$	$17.0$	$32.7$
Loco $\to$ Kick	in motion, $25\%$ range	$97.2$	$90.2$	$14.7$	$29.8$
Loco $\to$ Kick	in motion, $50\%$ range	$97.0$	$87.8$	$13.7$	$27.5$
Loco $\to$ Kick	in motion, full range	$95.8$	$84.1$	$11.9$	$25.7$

TABLE VIII : Handoff robustness, both directions.

	$v_{x}$ MAE		$v_{y}$	$\omega_{z}$	Terr.	Push
	fwd.	back.	(m/s), (rad/s)		(%)	(%)
Loco-only, seed $42$	$0.39$	$0.51$	$0.35$	$0.21$	$98.7$	$100.0$
s.d. over $3$ seeds	$0.07$	$0.05$	$0.07$	$0.17$	$20.4$	$0.0$
With library	$0.42$	$0.55$	$0.35$	$0.25$	$100.0$	$100.0$
$\Delta$	$+0.02$	$+0.04$	$+0.00$	$+0.04$	$+1.3$	$+0.0$

TABLE XI : Locomotion generality with the skill library attached. Both policies build on the seed-

42

substrate, and

\Delta

is that paired difference, computed before rounding. The s.d. row is the spread of three Stage-1 seeds (

42/7/13

), a noise floor rather than an interval around the row above: their terrain rates are

98.7/97.3/62.7

Skill	$1$	$2$	$3$	$4$	$5$	$6$	$7$	All
$\delta$ (deg)	$+21$	$-40$	$-8$	$-11$	$-5$	$-17$	$0$	–
Suc.@ $0.5$ (%)	$10.4$	$1.2$	$21.8$	$26.8$	$36.2$	$8.6$	$9.6$	$16.4$
+ calibration	$\mathbf{28.8}$	$\mathbf{20.2}$	$\mathbf{29.0}$	$\mathbf{41.6}$	$\mathbf{40.4}$	$\mathbf{27.4}$	$\mathbf{16.8}$	$\mathbf{29.2}$
Suc.@ $1.0$ (%)	$22.4$	$4.8$	$38.8$	$52.2$	$65.2$	$24.0$	$20.2$	$32.5$
+ calibration	$\mathbf{56.0}$	$\mathbf{42.4}$	$\mathbf{49.0}$	$\mathbf{71.6}$	$\mathbf{74.8}$	$\mathbf{51.2}$	$\mathbf{33.6}$	$\mathbf{54.1}$

TABLE XII : Per-skill angular calibration, medians over Table VI ’s

16

checkpoints, on a separate

250

-trial scan.

\delta

is fitted on half of each checkpoint’s trials and scored on the other, then swapped.

Explore similar work

SkillX: Unified Multi-Skill Policy Learning for Humanoid Soccer
Date pendingZhangchen Ye, Enxuan Ruan, Yifei Bao +10Badminton Human-To-Robot Transfer
Noetix Robotics · Tsinghua University
RoboNaldo: Accurate, Stable and Powerful Humanoid Soccer Shooting via Motion-Guided Curriculum Reinforcement Learning
Jun 9, 2026Yichao Zhong, Yidan Lu, Yuhang Lu +9Badminton Full-Body Humanoid Control
The University of Hong Kong · Archon Robotics · The Chinese University of Hong Kong
Light-Loco-Parkour: Versatile Perceptive Whole-Body Locomotion via Multi-Skill Distillation
Aug 1, 2026Hongming Chen, Zhuoran Li, Hongxi Wang +9Whole-Body Control Locomotion
Light Origins · Light Origins robotics team Co-Lead

Date pending, cs.ROJ/K move · Enter open · S save

SkillX: Unified Multi-Skill Policy Learning for Humanoid Soccer

Zhangchen Ye, Enxuan Ruan, Yifei Bao, Runhan Huang+9

Noetix Robotics · Tsinghua University

Badminton Human-To-Robot Transfer Whole-Body Control Skills

Open paper AskPDF arXiv

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.

Abstract

Explore similar work

SkillX: Unified Multi-Skill Policy Learning for Humanoid Soccer

RoboNaldo: Accurate, Stable and Powerful Humanoid Soccer Shooting via Motion-Guided Curriculum Reinforcement Learning

Light-Loco-Parkour: Versatile Perceptive Whole-Body Locomotion via Multi-Skill Distillation