Learning to Brachiate via Simplified Model Imitation
Event Type
Technical Paper
Interest Areas
Research & Education
Presentation Types
Registration Categories
Full Conference Supporter
Full Conference
Virtual Conference Supporter
Virtual Conference
Exhibitor Additional Full Conference
Exhibitor Full Conference
DescriptionWe develop highly dynamic 2D brachiation skills for a gibbon model using reinforcement learning. A control policy for a simplified model is learned, which then guides the learning for a 14-link articulated model. Our results demonstrate brachiation motions with significant flight phases and emergent swing-pumping behavior.