Recent robotic manipulation competitions have highlighted that sophisticated robots still struggle to achieve fast and reliable perception of task-relevant objects in complex, realistic scenarios. To improve these systems' perceptive speed and robustness, we present SegICP, a novel integrated solution to object recognition and pose estimation. SegICP couples convolutional neural networks and multi-hypothesis point cloud registration to achieve both robust pixel-wise semantic segmentation as well as accurate and real-time 6-DOF pose estimation for relevant objects.Our architecture achieves 1 cm position error and < 5 • angle error in real time without an initial seed. We evaluate and benchmark SegICP against an annotated dataset generated by motion capture.

Convolutional neural networks (CNNs), particularly those designed for object segmentation and pose estimation, are now applied to robotics applications involving mobile manipulation. For these robotic applications to be successful, robust and accurate performance from the CNNs is critical. Therefore, in order to develop an understanding of CNN performance, several CNN architectures are benchmarked on a set of metrics for object segmentation and pose estimation. This thesis presents these benchmarking results, which show that metric performance is dependent on the complexity of network architectures. The reasons behind poor pose estimation and object segmentation, which include object symmetry and resolution loss due to downsampling followed by upsampling in the networks, respectively, are also identified in this thesis. These findings can be used to guide and improve the development of CNNs for object segmentation and pose estimation in the future