Package information

panda-autograsp is an autonomous ROS based grasping solution that works with the Panda Emika Franka robot. In this grasping solution, several opensource grasping solutions are implemented on the Panda Emika Franka robot robot. The panda-autograsp package currently contains the following grasping algorithm:

• BerkleyAutomation/gqcnn

Note

These solutions work both on a physical as well as a simulated version of the panda robot. A simulated version of the panda robot is shipped with this package.

An video showing the panda-autograsp algorithm in action.

Package overview

Flow diagram of the panda-autograsp algorithm.

• Image processing nodes (Iai_kinect2_bridge): The image processing is performed by the iai_kinect2 package.
• panda_autograsp_server: This node is responsible for connecting all of the individual components of the panda-autograsp solution together.
• panda_autograsp_cli: This node is used to control the panda-autograsp solution.
• grasp_planner_server: This node computes a valid grasp out of RGB-D images it receives from the panda-autograsp_server node.
• tf2_broadcaster: This node sends the Robot eye-hand calibration results to the /tf node so that the scene gets updated.
• moveit_planner_server: This node is used to control the robot using the moveit planning framework.

Grasping solutions

GQ-CNN & F-GQ-CNN

GQ-CNNs are neural network architectures that take as input a depth image and grasp, and output the predicted probability that the grasp will successfully hold the object while lifting, transporting, and shaking the object.

Original GQ-CNN architecture from Dex-Net 2.0.

Alternate faster GQ-CNN architecture from FC-GQ-CNN.

The GQ-CNN weights were trained on datasets of synthetic point clouds, parallel jaw grasps, and grasp metrics generated from physics-based models with domain randomization for sim-to-real transfer. See the ongoing Dexterity Network (Dex-Net) project for more information.

Note

Currently, only the parallel jaw variants of the GQ-CNN and FC-GQ-CNN networks are supported by the panda-autograsp package. As a result, for the GQ-CNN’s and FC-GQ-CNN, the following network models can be chosen:

• GQCNN-2.0: Deep Learning to Plan Robust Grasps with Synthetic Point Clouds and Analytic Grasp Metrics. Trained on Dex-Net 2.0.
• GQCNN-2.1: Extension off GQ-CNN 2.0. The network that was trained on Dex-Net 2.0 is improved using RL in simulations.
• GQCNN-4.0-PJ: Improvement of the GQCNN-2.0 which computes more accurate grasps. This network module does additionally state whether it is better to grasp an object using suction or a parallel jaw gripper.
• FC-GQCNN-4.0-PJ: Modification of GQCNN-4.0-PJ in which a fully connected grasp quality CNN (FC-GQ-CNN) is used. This model has a faster grasp computation time and a more accurate grasp.

You can switch between these networks by supplying the panda-autograsp launch file with the model_type:=<MODEL_NAME> argument.