Organizations: Shanghai Jiao Tong University · Shanghai Artificial Intelligence Laboratory · Shanghai Innovation Institute · Huazhong University of Science and Technology · University of Science and Technology of China
Achieving human-level competitive intelligence and physical agility in humanoid robots remains a profound challenge, particularly in contact-rich and highly dynamic tasks such as boxing. While Multi-Agent Reinforcement Learning offers a principled framework for strategic interaction, its direct application to unstructured raw motor spaces inevitably leads to joint-level physical collapse, preventing the emergence of any viable combat tactics. To resolve this fundamental conflict between strategic exploration and physical feasibility, we formulate the humanoid combat task as a novel two-player latent-space zero-sum Markov game. Under standard regularity and approximate best-response assumptions, we show that the latent formulation induces an equivalent game over the decoder-reachable action manifold, providing an approximate-Nash interpretation of the resulting self-play dynamics. To instantiate this theoretical formulation, we propose RoboStriker, a hierarchical framework that decouples high-level reasoning from low-level execution. It first distills the tracking expertise of predefined boxing motions into a topologically bounded latent manifold. This structured latent foundation subsequently drives multi-agent co-evolution via Latent-Space Neural Fictitious Self-Play. Extensive experimental results demonstrate that gaming within this structured latent space substantially outperforms direct exploration. By constraining strategic exploration through a pretrained motion decoder, RoboStriker substantially reduces the catastrophic balance failures observed in raw action-space methods and achieves superior tactical performance in both competitive win rates and striking efficiency. Finally, we successfully deploy and validate our learned combat policies on real-world humanoid robots. Our code and video and supplementary materials are available at RoboStriker.
Figures & tables
Fig. 1 : Overview of RoboStriker. Stage I pretrains a motion tracker to produce physically plausible humanoid behaviors; Stage II compresses these behaviors into a bounded latent space for high-level control; Stage III(a) runs warm-up training on top of Stage II, then followed with Stage III(b) , a NFSP over latent-space to solve the humanoid boxing as a two-player latent space zero-sum game.
Methods
Win Rate (%) vs. Opponent
ηhit↑
ER ↑
(1)
(2)
(3)
(4)
(5)
(6)
(7)
(8)
(9)
(1) RoboStriker
–
68.52
76.24
82.41
84.47
95.38
98.50
100.0
92.50
0.685 ± 0.03
0.824 ± 0.02
(2) Fictitious SP
–
–
62.35
71.84
75.16
92.11
96.18
99.42
88.14
0.420 ± 0.03
0.650 ± 0.04
(3) Naive SP
–
–
–
68.45
70.38
88.61
94.52
98.53
85.25
0.350 ± 0.04
0.580 ± 0.05
(4) RoboStriker w/o AMP
–
–
–
–
62.12
82.14
88.37
95.18
78.40
0.490 ± 0.03
0.720 ± 0.05
(5) PPO-Only
–
–
–
–
–
45.24
77.86
82.11
68.55
0.231 ± 0.03
0.495 ± 0.02
TABLE I : Win Rate (%) of Cross-Playing and Tactical Performance. Each Win Rate is the win rate of the row agent against the column agent, averaged over 20 evaluation rounds with 4,096 parallel bouts per round. Offensive Landing Rate ( ηhit ) and Engagement Rate (ER) are evaluated against Naive Self-Play (Latent). Bold values indicate the top-1 values in the table.
Fig. 2 : Comprehensive performance evaluation. Subfigures (a) and (b) show the training convergence of base stability and offensive landing rate across all methods. Subfigures (c) and (d) compare RoboStriker with direct 29-DoF action-space self-play with and without AMP, isolating the effects of motion-prior regularization and latent-space action abstraction on tactical proficiency and physical stability. Shaded regions denote the standard deviation across five independent random seeds.
Fig. 3 : Real-world validation of RoboStriker.
Fig. 4 : Mujoco visualization of RoboStriker and the one without AMP.
Latent Dim.
MPKPE (mm)
Win Rate vs. 32
ηhit
8
128.2±4.4
7.1%
0.328±0.03
16
85.9±1.7
23.6%
0.334±0.03
32 (Ours)
38.5±0.4
–
0.685±0.03
64
35.0±0.4
39.2%
0.520±0.08
TABLE II : Ablation study on the latent dimension.
HoloBoxing presents a simple recipe for learning humanoid boxing by building on a pretrained whole-body controller. The recipe separates reusable motor skills from task decisions: a compact planner requests future body-point targets, while the frozen controller handles joint-level execution. A learned decoder maps 32-dimensional actions to structured target windows, and measured-state imitation initializes the planner before reinforcement learning. Task learning then refines interaction behavior while reusing the decoder and controller, avoiding the need to relearn whole-body control. This combination keeps task optimization focused on a compact upper policy and supplies an initial behavior before reward-driven exploration. Starting from that available controller, the recipe accounts for about seven minutes of additional training, including reused representation fitting and a single-GPU task run. This excludes controller pretraining, data preparation, evaluation, and search. Simulated boxing experiments and component controls evaluate the resulting behavior and the roles of imitation initialization and the target representation.
Competitive humanoid table tennis requires two distinct abilities: executing a desired return and deciding where to return the ball. We separate these two problems, with the ball landing target serving as an interface between motor control and strategy. (1) We first train a whole-body motor policy conditioned on the desired landing target, which maps the target to an executable humanoid stroke through a model-based strike planner and a learned controller. (2) We then improve the motor policy through self-play in closed-loop rallies, where the robot learns to handle incoming balls generated by another player rather than only prescribed incoming balls. At this stage, the landing targets are still sampled without considering the opponent, so the policy improves its execution without learning game strategy. (3) Finally, we freeze the entire motor stack and train a stroke-level strategic policy to choose landing targets through competitive self-play. Motor self-play improves both landing accuracy and legal-return rate during rallies. More importantly, with the entire motor stack fixed, changing only the target-selection policy leads to large differences in game outcomes: the learned strategy wins 89%, 82%, and 73% of points against center, random, and away-from-opponent baselines, respectively. The learned strategy also changes where it sends the ball according to the opponent's position and the previous target, showing that self-play learns more than simply aiming at fixed "good" locations. On hardware, the same frozen execution stack and strategic interface support complete 11-point games against a recreational human opponent, including multi-stroke rallies with both forehand and backhand returns.
Fanrong Dong, Ziyi Sun, Ziqi Han +6
Zhejiang University. · RoboParty Lab. · Ocean University of China. +3
High-speed racket sports provide a demanding testbed for humanoid robots, requiring time-critical decisions, precise striking, and dynamic whole-body coordination. In badminton, fast-changing shuttle trajectories require timely contact decisions, while successful returns demand precise racket pose and velocity within a brief contact window and across a broad three-dimensional striking workspace. Human motion data provide valuable priors for such athletic skills, but usable badminton references are limited and imperfect. Direct tracking provides insufficient executable variation for diverse shuttle conditions, while purely task-driven optimization may produce unnatural motion. We therefore present a three-stage hierarchical reinforcement learning framework for dynamic humanoid badminton. First, task-randomized motion augmentation expands sparse annotated hitting events into executable target-conditioned stroke variations, forming a continuous latent skill space. Second, a high-level planner outputs continuous latent skill codes to compose these skills online according to the observed shuttle state. Third, a context-conditioned adversarial regularizer encourages more natural planner-level skill usage while preserving return performance. In real-world deployment, our system sustains multi-skill rallies with human players, including forehand, backhand, and highly dynamic jump returns. This is the first real-world humanoid racket-sport system to demonstrate multi-skill human--robot rallies including highly dynamic jump returns.
Jingzhi Cui, Zhexiong Wang, Bangjie Xu +9
Tsinghua University · Hong Kong Embodied AI Lab · Zhejiang University +3