Mihir Kulkarni

Developed a pipeline to optimally Grasp objects of variable shape, size, and orientation with a robotic arm using vision capabilities. Implemented VGG16 and ResNet50 based network architecture using Transfer Learning in PyTorch. Implemented the pipeline on a custom 3D printed 5-DoF robotic arm using MoveIt! and KinectV2 depth camera. The problem of robotic grasping is still an unsolved problem with many approaches trying to generalize grasp predictions for unseen and dynamic en- vironments. Here we explore two approaches, one based on transfer learning, and another using a popular grasp detection model known as GG-CNN. In the transfer learning approach we tried 2 base models, VGG-16 and ResNet- 50. ResNet-50 provided better results with a testing accuracy of 83.3% while VGG-16 provided an accuracy of 78.2%. In order to test our model on a real robotic arm, we built a 5-DOF arm and added a custom parallel plate gripper.
Complete ROS and Moveit support is added to our developed robotic arm. The processed RG-D image from the KinectV2 camera is given as an input to the model which predicts the 5-D grasp configuration. Required electronic system design and its PCB is built which controls the robotic arm. The pre- dicted 5-D grasp configuration is then transformed to the object pose w.r.t the base link frame of the robot. Finally, A ROS node that automates the task of picking objects lying in different positions & orientations and sends the joint angle values over pyserial communication to the Arduino (PCB) is written. Thus, we have developed a complete pipeline for the task of Deep Learning based robotic grasping.