Peg-in-hole insertion is a task where the robot attempts to position the end-effector at a given socket. Conventional setups have a camera aligned with the end-effector to get an unobstructed view of the socket. In our specific setup for space satellite service, the Mission Robotic Vehicle (MRV) attempts to insert a payload placed on a robot arm into the socket on the client satellite with visual feedback from a camera placed on the base satellite. Due to the nature of this setup, the robot arm can block some part of the visual feedback and the socket may not be directly visible. In this work, we develop a simulated environment that best reproduces the real world environment and propose two methods of dealing with occlusion in this scenario based on computation resources. The first method uses knowledge of the pose of the robot arm to estimate occluded areas. The projection of the arm onto the image plane forms a mask of occluded regions and the tracker will avoid masked regions. The second method is similar to the previous method except the mask is generated from learning-based image segmentation. Both methods are robust with various levels of occlusion. The first method requires very low computational power. The second method can function without knowledge of arm specifications and poses and provides a better estimate of occlusion. Our observation shows that both methods improve tracking accuracy and reduce the rate at which the tracker loses track due to occlusion.

48 pages

Thesis Committee:
Howie Choset (Chair)
Matthew J. Travers

Srinivasan Seshan, Head, Computer Science Department
Martial Hebert, Dean, School of Computer Science

Return to: SCS Technical Report Collection
School of Computer Science

This page maintained by reports@cs.cmu.edu