Extended residual learning with one-shot imitation learning for robotic assembly in semi-structured environment

Area of Science:

• Robotics
• Machine Learning
• Computer Vision

Background:

• Robotic assembly demands precise manipulation and coordination, often requiring advanced learning techniques.
• Existing methods like residual reinforcement learning face limitations in semi-structured environments due to reliance on hand-designed features or extensive demonstrations.

Purpose of the Study:

• To propose an innovative Object-Embodiment-Centric Imitation and Residual Reinforcement Learning (OEC-IRRL) approach for robotic assembly.
• To enhance learning efficiency and effectiveness in semi-structured environments by minimizing environmental interactions and utilizing a single demonstration.

Main Methods:

• Developed an object-embodiment-centric (OEC) task representation integrating vision models with imitation and residual learning.
• Employed imitation learning with via-point movement primitives for a generalizable base policy.
• Utilized residual reinforcement learning for uncertainty-aware policy refinement during assembly.

Main Results:

• The OEC task representation positively impacted base and residual policy learning in semi-structured settings.
• Achieved a 46% improvement in success rates and a 25% reduction in assembly time.
• Required only a single demonstration and less than 1.2 hours of interaction.

Conclusions:

• The OEC-IRRL approach offers a viable solution for robotic assembly tasks.
• The method eliminates the need for specialized expertise or custom fixtures, broadening applicability.