Iterative transfer learning for automatic collective motion tuning on multiple robot platforms.

Area of Science:

• Robotics
• Artificial Intelligence
• Machine Learning

Background:

• Achieving coordinated collective motion in multi-robot systems is challenging.
• Traditional methods require extensive data collection and hardware testing.
• Transfer learning offers a potential solution for efficient behavior adaptation.

Purpose of the Study:

• To propose and evaluate an iterative transfer learning approach for swarming collective motion.
• To enable rapid tuning of stable collective behaviors across diverse robot platforms.
• To reduce the need for large datasets and minimize hardware-based trial-and-error.

Main Methods:

• An iterative transfer learning framework was developed.
• A deep learner was trained to recognize swarming collective motion.
• The system utilized minimal initial training data from each robot platform.
• The approach was validated on simulated Pioneer 3DX and real Sphero BOLT robots.

Main Results:

• The transfer learning approach successfully enabled automatic tuning of stable collective behaviors on both simulated and real robots.
• The iterative tuning procedure proved to be fast and accurate.
• The learned behaviors were transferable to multi-robot tasks like coverage.

Conclusions:

• Iterative transfer learning is an effective method for achieving swarming collective motion in mobile robots.
• This approach significantly reduces data collection costs and hardware risks.
• The method demonstrates adaptability for various multi-robot applications.