MetaGrasp: Generalizable Dexterous Multifingered Functional Grasping With Gradual Skill Curriculum Learning.

Area of Science:

• Robotics
• Artificial Intelligence
• Machine Learning

Background:

• Dexterous manipulation with multifingered robotic hands is complex due to high degrees of freedom.
• Existing methods struggle with task-specific adaptations and generalizing to new grasping tasks or poses.

Purpose of the Study:

• To introduce MetaGrasp, a novel approach for dexterous functional grasping using multitask reinforcement learning.
• To develop a versatile and adaptive grasping policy capable of learning from object point clouds and grasp pose classifications.

Main Methods:

• Defined dexterous functional grasping as a multitask reinforcement learning problem based on hand grasp pose classification.
• Implemented a gradual skill curriculum learning (GSCL) framework with three difficulty stages: beginner, intermediate, and advanced.
• Combined meta imitation learning (IL) with curriculum learning for a five-fingered robotic hand.

Main Results:

• MetaGrasp demonstrated superior one-shot generalization ability on new grasp tasks compared to existing methods.
• The trained policy adapted to grasp diverse object instances and categories based on functional grasp intentions from expert demonstrations.
• Outperformed state-of-the-art single-task dexterous grasping methods in experimental evaluations.

Conclusions:

• MetaGrasp offers a robust and adaptive solution for dexterous robotic grasping.
• The multitask RL framework with GSCL significantly improves generalization and efficiency in robotic manipulation.
• This approach enables robotic hands to perform precise functional grasps with minimal system interaction.