Spherical Latent Motion Prior for Physics-Based Simulated Humanoid Control

Jing Tan1,*, Weisheng Xu1,*, Xiangrui Jiang1,*, Jiaxi Zhang1, Kun Yang1, Kai Wu1, Jiaqi Xiong2, Shiting Chen1, Yangfan Li1, Yixiao Feng1, Yuetong Fang1, Yujia Zou1, Yiqun Song1, Renjing Xu1,†
1The Hong Kong University of Science and Technology (Guangzhou) 2University of Oxford

*Equal contribution    Corresponding author

Abstract

Learning motion priors for physics-based humanoid control is an active research topic. Existing approaches mainly include variational autoencoders (VAE) and adversarial motion priors (AMP). VAE introduces information loss, and random latent sampling may sometimes produce invalid behaviors. AMP suffers from mode collapse and struggles to capture diverse motion skills. We present the Spherical Latent Motion Prior (SLMP), a two-stage method for learning motion priors. In the first stage, we train a high-quality motion tracking controller. In the second stage, we distill the tracking controller into a spherical latent space. A combination of distillation, a discriminator, and a discriminator-guided local semantic consistency constraint shapes a structured latent action space, allowing stable random sampling without information loss. To evaluate SLMP, we collect a two-hour human combat motion capture dataset and show that SLMP preserves fine motion detail without information loss, and random sampling yields semantically valid and stable behaviors. When applied to a two-agent physics-based combat task, SLMP produces human-like and physically plausible combat behaviors only using simple rule-based rewards. Furthermore, SLMP generalizes across different humanoid robot morphologies, demonstrating its transferability beyond a single simulated avatar.

Method Overview

SLMP method overview

Video Results

Random Sampling Comparison with Prior Methods

ASE

PULSE

SLMP

Latent Representation Comparison

VAE

VQ-VAE

Sphere

SLMP

Loss Ablation: Random Sampling Stability

Distill only

Distill + GAN

Distill + NSC

SLMP (Full)

Two-Agent Combat: Reward Design Comparison

NCP Reward

NCP + AMP Reward

Simple Rule-Based Reward

Realistic Robot Validation – Unitree G1

Random Sampling

Combat

Realistic Robot Validation – ENGINEAI PM01

Random Sampling

Combat

BibTeX

@article{tan2026slmp,
  title   = {Spherical Latent Motion Prior for Physics-Based Simulated Humanoid Control},
  author  = {Tan, Jing and Xu, Weisheng and Jiang, Xiangrui and Zhang, Jiaxi and Yang, Kun and Wu, Kai and Xiong, Jiaqi and Chen, Shiting and Li, Yangfan and Feng, Yixiao and Fang, Yuetong and Zou, Yujia and Song, Yiqun and Xu, Renjing},
  journal = {arXiv preprint arXiv:2603.01294},
  year    = {2026}
}