Related Experiment Video
Updated: Nov 26, 2025
10:19
Studying the Neural Basis of Adaptive Locomotor Behavior in Insects
Published on: April 13, 2011
13.1K
Multi-expert learning of adaptive legged locomotion
Chuanyu Yang1, Kai Yuan1, Qiuguo Zhu2
1School of Informatics, University of Edinburgh, Edinburgh, UK.
Science Robotics
|December 10, 2020
Summary
This study introduces a multi-expert learning architecture (MELA) for robots to adapt locomotion skills. MELA dynamically synthesizes new motor skills, enabling versatile and responsive robot movement in unseen situations.
Area of Science:
- Robotics
- Machine Learning
- Artificial Intelligence
Background:
- Versatile robot locomotion demands adaptive motor skills for novel environments.
- Current methods struggle with real-time adaptation to unforeseen situations.
Purpose of the Study:
- To develop a multi-expert learning architecture (MELA) for generating adaptive robot locomotion skills.
- To enable robots to dynamically synthesize new motor skills for unseen scenarios.
Main Methods:
- MELA utilizes a gating neural network (GNN) to combine pre-trained deep neural networks (DNNs) representing expert skills.
- During runtime, MELA dynamically synthesizes DNNs to create adaptive behaviors.
- The framework leverages existing expert skills and online policy synthesis.
Main Results:
- Demonstrated successful multi-skill locomotion on a quadruped robot using a unified MELA framework.
- The robot autonomously performed trotting, steering, and fall recovery.
- MELA generated adaptive behaviors for changing tasks and unseen scenarios.
Conclusions:
- MELA effectively generates adaptive and versatile robot locomotion skills.
- The multi-expert learning approach enhances a robot's ability to respond to dynamic environments.
- This framework shows significant potential for robust autonomous robot navigation.
