Robotic Grasping Detection with PointNet
WHERE TO GRAB ON EVERYDAY OBJECTS
Humans quickly learn how to hold objects correctly, like grabbing a toothbrush by the handle. This skill develops early in life, with babies mastering the "pincer grasp" by around 1 year old.
In robotics, there's a lot of research on teaching robots to grasp objects. This project isn't about physically grasping objects but figuring out where to grasp them. We're training robots to recognize the optimal points to grasp on different objects, similar to how humans learn to find handles.
This project involves creating a PointNet model from scratch. We collected point cloud data using an iPhone's LiDAR via the Polycam app, focusing on three objects: Cup, Knife, and Pan. To clean our data, we used the RANSAC algorithm to remove any outliers. Due to our limited dataset, traditional image data augmentation techniques like scaling or reflection weren't suitable for point cloud data. We then labeled our point cloud using Segments.ai and processed the segmentation labels. Finally, we trained our model for both classification and part-segmentation, aiming to segment the handle on the object. While the model didn't perform as well on an out-of-sample dataset, it achieved over 98% accuracy on our test set.
