Cloud-based Robot Grasping with the Google Object Recognition Engine

Cloud-based Robot Grasping with the Google Object Recognition Engine

Motivation:

For today’s robots,  the constraints on power and computing speed is a big problem. In this paper, the author wants to show a new cloud-based grasping system which can offload some computing process to the cloud and utilize the efficiency of cloud and Internet to achieve itself. The grasping system is used on robots which can grasp common household objects.

Background:

Object recognition is very well-studied topic in computer science. Researchers are working to improve both the scalability and accuracy of large-scale image recognition, making object recognition systems commercially viable. The cloud has the ability to enable information sharing between networked robots to accomplish tasks widely separated in time and space.

Main Points:

1. Although the network based robots have a long history, the cloud computing is a new powerful paradigm for massively parallel computation and real-time sharing of vast data resources. Cloud robots have the potential to improve indexing vast libraries, massively-parallel sampled-based motion planning, sharing of outcomes and obtaining on-demand human guidance when needed.
2. For the system architecture, it consists of two parts, the cloud part and the robot part. In the robot part, it has a camera, a 3D sensor, a pose estimation and selected feasible grasp. In the cloud part, it has Google object recognition engine, Google cloud storage and grasp analysis. The system also has two phases, the online phase and the offline phase.
3. The whole recognizing and grasping process consists of training of the object recognition server , the creation and analysis of the candidate grasp and the pose estimation and grasping.
4. In the experiments done by the author, he experimented with set of six household objects. For some objects such as the air freshener and mustard bottle, small errors in pose estimation had a significant effect on grasp outcome.

Trade-offs:

1. The whole system incorporated many small systems such as the Google Goggles, GraspIt!, ASUS Xtion PRO sensor and make them work well. Maybe the future can be based on a single system which can get all the strength of these existing systems and use more compatible ports to connect each other.
2. As talked in the paper, the author can experiment with larger object sets in the future. Because the situation is based in home and the robot may need to grasp some bigger things than the juice bottle or the soap.
3. The author only mentions that there are two phases for the system while he did not compare the efficiency of these two. I’m interested in how much could the online phase improve the efficiency than the normal offline phase.