DexNet

DexNet is a research project and a software suite developed at the University of California, Berkeley, focused on robust grasping and manipulation by robots. Its primary goal is to provide robots with the ability to reliably grasp objects in unstructured environments with minimal human intervention.

DexNet utilizes a data-driven approach, employing a large-scale dataset of 3D object models and simulated grasp outcomes. The core of the DexNet system lies in its grasp quality metrics, which quantify the robustness and stability of a grasp based on factors such as force closure, wrench resistance, and the potential for slippage.

The software suite provides functionalities for:

• Grasp Quality Evaluation: Analyzing the quality of a potential grasp based on the object's geometry and the robot's gripper configuration. This involves calculating various grasp quality metrics.
• Grasp Planning: Generating a set of candidate grasps for an object based on its 3D model. This process often involves searching through a large space of possible grasp configurations and selecting the most promising ones based on the grasp quality metrics.
• Grasp Execution: Providing instructions to a robot to execute a selected grasp. While the DexNet software itself does not directly control the robot's motors, it provides the necessary information for a robot controller to perform the grasp.
• Robotic Perception: Using computer vision techniques to estimate the 3D shape and pose of objects in the environment. This information is then used for grasp planning.

The DexNet project has explored different gripper types, including parallel-jaw grippers and suction grippers. It has also investigated various methods for dealing with uncertainty in object pose and shape, such as using tactile sensing or force feedback to refine the grasp during execution. The datasets and tools provided by DexNet are intended to accelerate research in robotic grasping and manipulation.