Imitation-relaxation reinforcement learning for sparse badminton strikes via dynamic trajectory generation

Area of Science:

• Robotics
• Artificial Intelligence
• Control Systems

Background:

• Robotic racket sports, like badminton, are challenging due to complex shuttlecock dynamics and reinforcement learning's sparse reward issues.
• Achieving precise dynamic motion control in robots for fast-paced sports is a significant hurdle.

Purpose of the Study:

• To develop a novel learning framework, DTG-IRRL, for improving robot performance in executing badminton strikes.
• To address the challenges of non-linear dynamics and sparse rewards in reinforcement learning for robotic badminton.

Main Methods:

• Integration of imitation-relaxation reinforcement learning with dynamic trajectory generation.
• Analysis of reward function convergence within a specific parameter space to understand learning difficulties.
• Zero-shot transfer implementation on hardware for real-world testing.

Main Results:

• The DTG-IRRL framework demonstrated significantly improved training efficiency and faster convergence.
• Achieved twice the landing accuracy compared to previous methods.
• Hardware implementation resulted in a 90% hitting rate and 70% landing accuracy, enabling sustained human-robot rallies.

Conclusions:

• The proposed DTG-IRRL framework effectively mitigates convergence issues caused by sparse rewards in robotic badminton.
• The study highlights the generalizability of the framework across different robotic platforms, like the UR5 robot.
• High dynamic performance of robotic arms is crucial for success in racket sports applications.