Deep Learning Method for Grasping Novel Objects Using Dexterous Hands.

Area of Science:

• Robotics
• Computer Vision
• Machine Learning

Background:

• Robotic grasping capabilities significantly lag behind human dexterity, presenting a major challenge in robotics research.
• Humans intuitively adapt grasping postures based on object features and hand/palm orientation, a skill not yet replicated in robots.

Purpose of the Study:

• To develop advanced grasping posture prediction networks (GPPNs) that leverage both object imagery and dexterous hand palm pose.
• To integrate GPPNs with grasping rectangle detection networks (GRDNs) to create multilevel convolutional neural networks (ML-CNNs) for enhanced robotic grasping.

Main Methods:

• Grasping posture prediction networks (GPPNs) were designed with multiple inputs, incorporating object images and dexterous hand palm poses.
• GPPNs were combined with grasping rectangle detection networks (GRDNs) to form multilevel convolutional neural networks (ML-CNNs).
• A force-closure index was used to evaluate grasping quality, with postures generated in the GraspIt! simulation environment. Datasets were created using depth images from the Gazebo environment.

Main Results:

• Simulation experiments in GraspIt! demonstrated that ML-CNNs achieve high grasping quality.
• Studies using a variable-controlling approach confirmed the influence of image and palm pose inputs on GPPN performance.
• Comparison with existing grasp detection methods showed superior performance of the ML-CNNs.

Conclusions:

• The developed ML-CNNs show significant promise for improving robotic grasping accuracy and performance.
• The proposed method accurately completes grasping tasks for novel objects on a physical robotic platform (Shadow hand).
• This research contributes to bridging the gap between human and robotic grasping abilities through advanced deep learning techniques.