Gripper-Agnostic Grasping
learned method for robotic grasping across grippers
In collaboration with Anthony Simeonov, this project explored ways to condition robotic manipulation of novel, cluttered objects on gripper shape and manipulation task. In the way that you wouldn’t grab a mug by its rim if you wanted to flip it upside down, we want to generate grasps that depend on (1) gripper reach and extents and (2) what the robot intends to do with the object. Our approach built off of the existing GraspNet architecture presented by Mousavain et al.
Preliminary results on the gripper-dependent part of this task were presented at the IPPC Workshop at IROS 2023, and then again at NERC 2023. Read the paper here, and see the more current version of the poster (NERC) here.
Using: PyTorch, ROS, PointNet++ architecture, GraspNet architecture, CVAEs, Docker, PyBullet, RViz