Self-organized Learning from Synthetic and Real-World Data for a Humanoid Exercise Robot

Area of Science:

• Robotics
• Artificial Intelligence
• Machine Learning

Background:

• Humanoid robots require lifelong learning capabilities to train diverse users in physical exercises.
• Existing methods struggle to adapt to individual user morphologies and prevent knowledge forgetting over time.

Purpose of the Study:

• To develop a neural learning approach for humanoid exercise robots capable of analyzing and correcting physical exercises.
• To enable robots to learn from and adapt to multiple human partners with varying body types.
• To ensure robots can retain learned exercise information without catastrophic forgetting.

Main Methods:

• A modified Grow-When-Required (GWR) network with recurrent connections and episodic memory was developed.
• A novel subnode mechanism was introduced for learning spatiotemporal relationships in body movements and poses.
• The network grows online with new demonstrations, storing and retrieving individual human exercise data via the subnode mechanism.

Main Results:

• The Subnode-GWR network successfully stores and retrieves exercise information for individual users, adapting to different body morphologies.
• The system provides feedback on both pose and motion velocity, enabling accurate exercise correction.
• The approach demonstrated effective incremental online learning and prevention of catastrophic forgetting in both synthetic and real-world datasets.

Conclusions:

• The proposed GWR-based architecture facilitates lifelong learning in humanoid exercise robots.
• The system enables robots to adapt learned exercise references to different body variations through incremental online learning.
• This research contributes to more engaging and effective long-term human-robot interactions in physical training scenarios.