Learning ambidextrous robot grasping policies.

Area of Science:

• Robotics
• Artificial Intelligence
• Machine Learning

Background:

• Universal picking (UP) is crucial for robotic automation in e-commerce and manufacturing but faces challenges due to object diversity and environmental uncertainty.
• Current robotic grasping methods struggle with reliability and speed when handling a wide variety of novel objects in cluttered environments.

Purpose of the Study:

• To develop an advanced robotic grasping system, Dexterity Network (Dex-Net) 4.0, capable of high-reliability universal picking.
• To explore the effectiveness of using heterogeneous grippers in an "ambidextrous" grasping approach to enhance robotic manipulation capabilities.

Main Methods:

• Dex-Net 4.0 utilizes deep learning policies trained on large-scale synthetic datasets generated with domain randomization.
• The system integrates analytic models of physics and geometry for training grasp policies on parallel-jaw and suction cup grippers.
• Policies were trained using 5 million synthetic depth images, grasp attempts, and associated rewards from 3D object heaps.

Main Results:

• The Dex-Net 4.0 policy demonstrated consistent success in clearing bins containing up to 25 novel objects.
• Achieved a reliability rate exceeding 95% for universal picking tasks on a physical robot.
• The system operates at a high pick rate of over 300 mean picks per hour.

Conclusions:

• Dex-Net 4.0 significantly advances universal picking capabilities by effectively integrating heterogeneous grippers and deep learning.
• The "ambidextrous" grasping strategy combined with robust policy learning offers a scalable solution for complex robotic order fulfillment and manipulation tasks.
• This research provides a pathway for more versatile and reliable robotic systems in various industrial and service applications.