Evolving a Behavioral Repertoire for a Walking Robot

Area of Science:

• Robotics
• Machine Learning
• Evolutionary Computation

Background:

• Legged robots require versatile walking capabilities for real-world tasks.
• Existing algorithms primarily focus on straight-line locomotion, limiting robot utility.
• A need exists for algorithms that enable multi-directional walking control.

Purpose of the Study:

• To introduce a novel evolutionary algorithm, TBR-Evolution, for discovering a repertoire of walking controllers.
• To enable legged robots to learn to walk in numerous directions simultaneously.
• To enhance the speed and efficiency of controller evolution by leveraging discarded solutions.

Main Methods:

• The Transferability-based Behavioral Repertoire Evolution (TBR-Evolution) algorithm was developed.
• Novelty search with local competition was employed to find diverse and high-performing solutions.
• A transferability approach combining simulation and real-world experiments was utilized.

Main Results:

• TBR-Evolution successfully discovered hundreds of walking controllers for various directions.
• The algorithm demonstrated significantly faster learning compared to independent controller evolution.
• A hexapod robot equipped with TBR-Evolution controllers could reach all points within its reachable space.

Conclusions:

• TBR-Evolution offers a new paradigm for simultaneously optimizing multiple robot behaviors.
• The algorithm enables efficient multi-directional locomotion for legged robots.
• This approach significantly expands the practical applicability of legged robots in complex environments.