Knowledge Distillation and Reinforcement Learning in a Human-Machine Collaboration Delivery System With a Robotic...

Area of Science:

• Robotics
• Artificial Intelligence
• Computer Vision

Background:

• Robotic arms are crucial for human-robot collaboration but struggle with dynamic environments and path planning.
• Conventional methods face limitations in adapting to real-world complexities and the simulation-to-reality gap.

Purpose of the Study:

• To explore robotic arm usability for object delivery in dynamic human environments.
• To develop advanced path planning methods overcoming traditional limitations using reinforcement learning.
• To enhance the transferability of robotic models by bridging the simulation-to-reality gap.

Main Methods:

• Employed reinforcement learning (RL) to train four path planning models: Approach RL, Delivery RL, Decision RL, and Merged Model.
• Utilized image segmentation to reduce discrepancies between simulated and real-world environments.
• Applied CycleGAN for image translation to transform real hand features into virtual ones, improving model transferability.

Main Results:

• The Decision RL Model achieved 99.17% accuracy, and the Merged Model reached 99.92% accuracy.
• The integrated approach demonstrated improved stability and accuracy in complex human-robot collaboration scenarios.
• Validated the effectiveness of combining RL, image segmentation, and CycleGAN for robotic arm applications.

Conclusions:

• The proposed method offers a scalable and efficient solution for robotic arms in dynamic domains.
• Reinforcement learning, image segmentation, and image translation effectively enhance robotic arm performance.
• This study confirms the feasibility of advanced AI techniques for robust human-robot interaction.