An Adaptive Imitation Learning Framework for Robotic Complex Contact-Rich Insertion Tasks.

Area of Science:

• Robotics
• Artificial Intelligence
• Machine Learning

Background:

• Complex contact-rich insertion is a key robotic skill, often requiring intricate trajectories and force control.
• Current hybrid learning frameworks demand numerous human demonstrations, even for similar tasks, increasing teaching burden.

Purpose of the Study:

• To develop an adaptive imitation framework that reduces human effort in teaching complex robotic insertion tasks.
• To enable learning of new tasks from a single demonstration within a class of topologically similar trajectories.

Main Methods:

• Introduced dynamic movement primitives into a hybrid trajectory and force learning framework.
• Utilized a single task instance's trajectory profile to learn a class of complex contact-rich insertion tasks.
• Employed imitation learning for trajectory generation and reinforcement learning for force control policies.

Main Results:

• The proposed framework demonstrated sample efficiency in learning complex insertion tasks.
• Experimental evaluations showed improved safety and generalization capabilities compared to traditional methods.
• Validated performance on both simulation environments and real robotic hardware.

Conclusions:

• The adaptive imitation framework significantly reduces human repetitive teaching efforts for robot manipulation.
• The approach offers a more efficient, safer, and generalizable method for learning complex contact-rich insertion tasks.
• This work advances robotic learning by enabling faster adaptation to new, yet similar, manipulation tasks.