PulseTrendingJournal ClubResearchersJournalsExplore
Instagram
HomeTrendingJournal ClubExplore
Synapse
⌘+K
Synapse
April 15, 2026ActuatorsOpen Access

HAML: Humanoid Adversarial Multi-Skill Learning via a Single Policy

View Full Paper
Ask AI
Bookmark
Share

Authors

XFXu FangHLHsin‐Yi LiaoYCYanyun Chen

Discussion

Loading...

Member takes

Overview

This approach demonstrates improved skill and transition coverage in humanoid controllers, indicating enhanced usability in real-world applications.

Key Points

  • The central aim is to develop a humanoid control system that efficiently translates motion datasets into a single-policy multi-skill framework.
  • Developed a two-stage learning system mapping motion datasets to a humanoid controller.
  • Utilized one-hot skill labels for automatic dataset construction with minimal manual effort.
  • Implemented a condition-aware loss to enhance controllability and reduce mode collapse.
  • Employed a teacher-student policy distillation strategy for robust real-world application.
  • Improved skill and transition coverage compared to previous approaches.
  • Enhanced realism and training efficiency in humanoid motion control.
  • Successfully operated on real hardware at 100 Hz with low latency of 15–25 ms.

Cite This Study

Fang et al. (2026) studied this question.

synapsesocial.com/papers/69df2c2fe4eeef8a2a6b131bhttps://doi.org/10.3390/act15040212
View Full Paper
Ask AI
Bookmark
Share