Digital-Twin-Assisted Skill Learning for 3C Assembly Tasks.

Area of Science:

• Robotics and Automation
• Artificial Intelligence
• Manufacturing Engineering

Background:

• Robotic automation in the 3C (computer, communication, consumer electronics) industry offers significant cost and efficiency benefits.
• Traditional programming and skill-learning methods struggle with complex, high-precision assembly tasks like flexible printed circuit (FPC) manipulation.

Purpose of the Study:

• To develop a novel learning-based framework for acquiring complex 3C assembly skills.
• To address the limitations of current methods in handling intricate robotic assembly tasks.

Main Methods:

• Construction of a multimodal digital-twin environment mirroring real-world 3C assembly processes.
• Collection of demonstration data using virtual reality (VR) devices, incorporating visual, tactile, and proprioception feedback.
• Development of a skill knowledge base through multimodal skill parsing to generate primitive policy sequences.
• Training of primitive policies using curriculum learning, residual reinforcement learning, and domain randomization.
• Transfer of learned skills from the digital-twin to the physical robotic system.

Main Results:

• The proposed framework successfully enables the acquisition of complex assembly skills.
• Demonstration of effective skill transfer from a simulated digital-twin environment to a real-world robotic system.
• Validation of the framework's effectiveness through experimental evaluation.

Conclusions:

• The learning-based framework, augmented by a multimodal digital-twin, provides an effective solution for complex 3C assembly tasks.
• This approach overcomes the limitations of manual programming and conventional skill learning for intricate robotic manipulations.
• The study validates the potential of digital twins and advanced machine learning techniques for advancing robotic assembly capabilities.