B.2: Dexnet

• Short Description: Dex-Net is a research project developed at UC Berkeley that focuses on improving the robustness and accuracy of robotic grasping. It involves the use of deep learning algorithms to train a robotic system to grasp a wide variety of objects in complex, cluttered environments. The system uses a dataset of 3D point cloud data, which is generated by scanning real-world objects, to train the grasping algorithms. Once trained, the system is able to generate grasping plans for new objects by predicting the poses at which they can be stably grasped. Additionally, Dex-Net uses a novel optimization algorithm to search for the best grasping poses, which allows it to generate high-quality grasps even in challenging situations. Overall, Dex-Net is a powerful tool for advancing the state of the art in robotic grasping.

• Reference, URL: https://berkeleyautomation.github.io/dex-net/

• Applicable Business Category:Manufacturing: WEEE & Battery

• Application in relevant Projects/Initiatives: Dex-Net research project

• Asset Type: Other, Code, datasets and algorithms

• AI Breadth: ML, Grasp Quality Convolutional Neural Networks (GQ-CNN)

• Learning Ability: Grasp Quality Convolutional Neural Networks (GQ-CNNs), Synthetic Data Generation

• Related Technologies: Python API

• Applicable Research Area: Collaborative AI,

• Applicable Technical Category: Robotics and automation

• License Information: Proprietary license

• Related to circularity and sustainability: Not directly

• Audience: Manufacturers, Service Robots

Go back to the AI portfolio of reference implementations