![Motivation
• Current methods in grasp learning for previously unseen objects
involve the transfer of features and primitives from humans to
robots.
• These transferred knowledge includes memory of previous
grasps[], human labelled datasets[], pre-defined visual/depth
features[]
• But these previous methods are inherently biased towards a
human’s knowledge and intuition
• For truly autonomous manipulators, we need a framework that
allows the robot itself to learn the knowledge and understanding
for grasping](https://image.slidesharecdn.com/graspslidespptx-150501034837-conversion-gate01/85/Grasp-slides-pptx-2-320.jpg)
Grasp slides pptx
- 1. Learning to grasp without human supervision
- 2. Motivation • Current methods in grasp learning for previously unseen objects involve the transfer of features and primitives from humans to robots. • These transferred knowledge includes memory of previous grasps[], human labelled datasets[], pre-defined visual/depth features[] • But these previous methods are inherently biased towards a human’s knowledge and intuition • For truly autonomous manipulators, we need a framework that allows the robot itself to learn the knowledge and understanding for grasping
- 3. Our Approach • We follow a learning by training approach where we have a dataset of images (and depth point clouds) of objects along with positive and negative grasp points • Since we need to eliminate the human in the loop, the robot itself should be able to collect training data • For this we have a robot execute random grasps on a wide variety of objects. The robot recognizes successful grasps using its inbuilt gripper force sensors
- 4. Grasp Data collected • Currently our robot is capable of executing about 80 random grasps an hour per arm. Which on a typical day(running for about 8 hours) yields about 80*8*2 = 1280 grasps a day. • As of now, the robot has executed about 17,000 random grasps and has around 1300 successful grasps. This is already larger than Cornell Grasp Dataset that is hand labelled and doesn't contain negative examples (failure grasps) • Associated with each grasp we record data from 2 RGB image cameras on the arm, 1 Depth camera (Creative Senz3D) and Kinect V2 attached to the head of the robot. The recorded streams include the entire motion of the arm in the grasp sequence