A model-free method to learn multiple skills in parallel on modular robots.

Area of Science:

• Robotics
• Control Systems
• Artificial Intelligence

Background:

• Legged robots excel in unstructured environments but need specialized control.
• The simulation-to-real gap hinders the direct application of simulated robot controllers.
• Rapid, model-agnostic real-world learning methods are crucial for robot deployment.

Purpose of the Study:

• To present a generic method for acquiring basic locomotion skills in legged robots.
• To address the challenge of transferring controllers from simulation to the real world.
• To enable fast, parallel learning of locomotion with minimal real-world trials.

Main Methods:

• Utilizing Central Pattern Generators (CPGs) for locomotion control.
• Focusing on optimizing initial states rather than connection weights within the CPG model.
• Employing mathematical analysis to underpin the controller model's novelty.
• Conducting empirical validation across six diverse robot morphologies.

Main Results:

• The proposed method enables robots to learn primary locomotion skills rapidly.
• Learning occurs in under 15 minutes in real-world experiments.
• The approach demonstrates effectiveness across various robot designs.
• Successful showcase of learned skills in a targeted locomotion task.

Conclusions:

• The CPG-based method effectively bridges the sim-to-real gap for legged robot locomotion.
• Optimizing initial states offers a viable alternative to weight optimization for CPGs.
• The technique facilitates quick, adaptable, and model-agnostic skill acquisition in real robots.