Functional Contour-following via Haptic Perception and Reinforcement Learning.

Area of Science:

• Robotics
• Artificial Intelligence
• Machine Learning

Background:

• Fine manipulation tasks often lack sufficient visual feedback, necessitating reliance on tactile and proprioceptive senses.
• Deformable, transparent, and visually occluded objects present significant challenges for robotic manipulation.

Purpose of the Study:

• To develop a real-time haptic perception and decision-making system for a robot performing a ziplock bag closure task.
• To investigate the efficacy of a Contextual Multi-Armed Bandit (C-MAB) reinforcement learning algorithm for efficient robotic learning.

Main Methods:

• A deep neural network classifier was trained to assess the zipper's state within a robotic grasp.
• A Contextual Multi-Armed Bandit (C-MAB) algorithm was employed to balance exploration and exploitation in the state-action space.
• The learned policy was validated on novel ziplock bag configurations and different contour-following tasks.

Main Results:

• The C-MAB learner demonstrated superior efficiency in exploring the state-action space compared to a benchmark Q-learner.
• The system successfully learned to perform the challenging ziplock bag closure task.
• The approach proved effective for contour-following tasks with materials like wire and rope.

Conclusions:

• This research advances reinforcement learning methods for robots operating with limited resources, such as hardware lifespan and research time.
• The developed haptic perception and decision-making approach enables efficient and effective robotic learning in physical testbeds for complex manipulation tasks.