Autonomous Functional Movements in a Tendon-Driven Limb via Limited Experience

Area of Science:

• Robotics
• Machine Learning
• Biologically-Inspired Systems

Background:

• Robots require efficient learning for widespread adoption, especially in complex, dynamic environments.
• Biological systems master multiple tasks with minimal trial-and-error despite intricate musculoskeletal structures.

Purpose of the Study:

• To demonstrate a few-shot autonomous learning method for robotic movement control.
• To enable robots to learn complex tasks with minimal attempts.

Main Methods:

• A model-free, open-loop approach termed G2P (General-to-Particular) was developed.
• The method involves initial motor babbling for an inverse map, followed by reinforcement learning and map refinement.

Main Results:

• Effective movement generation was achieved in simulation and hardware for a 3-tendon, 2-joint limb.
• The G2P algorithm facilitated quick, robust, and versatile adaptation.

Conclusions:

• The G2P algorithm offers a biologically-plausible pathway for rapid robotic learning.
• This approach has implications for both advanced robotics and understanding biological motor control.