Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

RoboStriker: Hierarchical Decision-Making for Autonomous Humanoid Boxing

Kangning Yin, Zhe Cao, Wentao Dong, Weishuai Zeng, Tianyi Zhang, Qiang Zhang, Jingbo Wang, Jiangmiao Pang, Ming Zhou, Weinan Zhang

TL;DR

RoboStriker tackles the challenge of autonomous humanoid boxing by decoupling high-level strategy from low-level motor control in a three-stage hierarchy. It learns a physically grounded motion tracker from motion capture, distills those motions into a compact latent space on the unit hypersphere $ \\mathbb{S}^{d-1}$, and trains competitive tactics in this latent space via Latent-Space Neural Fictitious Self-Play (LS-NFSP) with an adversarial motion prior and behavioral warmup. The approach yields stable training, superior cross-play performance, and sim-to-real transfer on a 29-DOF humanoid, outperforming baselines across tactical, stability, and stylistic metrics. By constraining strategy search and leveraging a topology-aware latent representation, the work provides a scalable pathway to embodied MARL for physically grounded robotic combat.

Abstract

Achieving human-level competitive intelligence and physical agility in humanoid robots remains a major challenge, particularly in contact-rich and highly dynamic tasks such as boxing. While Multi-Agent Reinforcement Learning (MARL) offers a principled framework for strategic interaction, its direct application to humanoid control is hindered by high-dimensional contact dynamics and the absence of strong physical motion priors. We propose RoboStriker, a hierarchical three-stage framework that enables fully autonomous humanoid boxing by decoupling high-level strategic reasoning from low-level physical execution. The framework first learns a comprehensive repertoire of boxing skills by training a single-agent motion tracker on human motion capture data. These skills are subsequently distilled into a structured latent manifold, regularized by projecting the Gaussian-parameterized distribution onto a unit hypersphere. This topological constraint effectively confines exploration to the subspace of physically plausible motions. In the final stage, we introduce Latent-Space Neural Fictitious Self-Play (LS-NFSP), where competing agents learn competitive tactics by interacting within the latent action space rather than the raw motor space, significantly stabilizing multi-agent training. Experimental results demonstrate that RoboStriker achieves superior competitive performance in simulation and exhibits sim-to-real transfer. Our website is available at RoboStriker.

RoboStriker: Hierarchical Decision-Making for Autonomous Humanoid Boxing

TL;DR

RoboStriker tackles the challenge of autonomous humanoid boxing by decoupling high-level strategy from low-level motor control in a three-stage hierarchy. It learns a physically grounded motion tracker from motion capture, distills those motions into a compact latent space on the unit hypersphere , and trains competitive tactics in this latent space via Latent-Space Neural Fictitious Self-Play (LS-NFSP) with an adversarial motion prior and behavioral warmup. The approach yields stable training, superior cross-play performance, and sim-to-real transfer on a 29-DOF humanoid, outperforming baselines across tactical, stability, and stylistic metrics. By constraining strategy search and leveraging a topology-aware latent representation, the work provides a scalable pathway to embodied MARL for physically grounded robotic combat.

Abstract

Achieving human-level competitive intelligence and physical agility in humanoid robots remains a major challenge, particularly in contact-rich and highly dynamic tasks such as boxing. While Multi-Agent Reinforcement Learning (MARL) offers a principled framework for strategic interaction, its direct application to humanoid control is hindered by high-dimensional contact dynamics and the absence of strong physical motion priors. We propose RoboStriker, a hierarchical three-stage framework that enables fully autonomous humanoid boxing by decoupling high-level strategic reasoning from low-level physical execution. The framework first learns a comprehensive repertoire of boxing skills by training a single-agent motion tracker on human motion capture data. These skills are subsequently distilled into a structured latent manifold, regularized by projecting the Gaussian-parameterized distribution onto a unit hypersphere. This topological constraint effectively confines exploration to the subspace of physically plausible motions. In the final stage, we introduce Latent-Space Neural Fictitious Self-Play (LS-NFSP), where competing agents learn competitive tactics by interacting within the latent action space rather than the raw motor space, significantly stabilizing multi-agent training. Experimental results demonstrate that RoboStriker achieves superior competitive performance in simulation and exhibits sim-to-real transfer. Our website is available at RoboStriker.
Paper Structure (65 sections, 1 theorem, 62 equations, 4 figures, 5 tables, 1 algorithm)

This paper contains 65 sections, 1 theorem, 62 equations, 4 figures, 5 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Physical Feasibility vs. Non-stationary Learning.
  3. Strategy Evolution vs. System Stability.
  4. Related Work
  5. Humanoid Motion Synthesis and Control.
  6. Skill-based Latent Representation Learning.
  7. Multi-Agent RL and Fictitious Self-Play.
  8. Method
  9. Two-player Zero-sum Markov Games
  10. Learning Expertise Motion Tracker
  11. Encoding Motion via Topological Latent Distillation
  12. Behavioral Warmup with Adversarial Priors
  13. Latent-Space Neural Fictitious Self-Play
  14. Theoretical Analysis
  15. Experiment
  16. ...and 50 more sections

Key Result

Proposition 3.5

Under Assumptions assump:bounded_strategy to assump:approx_br, the induced latent interaction admits the standard regularity conditions under which FSP dynamics in continuous zero-sum game are known to approach approximate Nash Equilibrium. Consequently, LS-NFSP will converge to an $\epsilon$-Nash

Figures (4)

  • Figure 1: Real-world clips of humanoid boxing using RoboStriker, showcasing agile, contact-rich punches and defenses under physical constraints.
  • Figure 2: 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 task as a two-player zero-sum game.
  • Figure 3: t-SNE visualization of the 32-dimensional latent manifold, illustrating the structured semantic clustering of combat primitives and their topological relationships that facilitate stable, compositional behavioral transitions in the hierarchical LS-NFSP framework.
  • Figure 4: Mujoco visualization of the LS-NFSP and the one without AMP.

Theorems & Definitions (4)

  • Definition 3.1: $\epsilon$-Nash Equilibrium aumann1976agreeing
  • Proposition 3.5
  • proof
  • proof