UC Berkeley

Several areas of computer science, including computer vision and natural

language processing, have witnessed rapid advances in performance, due in part

to shared datasets and benchmarks. In a robotics setting, benchmarking is

challenging due to the amount of variation in common applications: researchers can use

different robots, different objects, different algorithms, different tasks, and

different environments.

Cloud robotics, in which a robot accesses computation and data over a network,

may help address the challenge of benchmarking in robotics. By standardizing

the interfaces in which robotic systems access and store data, we can define a

common set of tasks and compare the performance of various systems.

In this dissertation, we examine two problem settings that are well served by

cloud robotics. We also discuss two datasets that facilitate benchmarking of

several problems in robotics. Finally, we discuss a framework for defining and

using cloud-based robotic services.

The first problem setting is object instance recognition. We present an

instance recognition system which uses a library of high-fidelity object models

of textured household objects. The system can handle occlusions, illumination

changes, multiple objects, and multiple instances of the same object.

The next problem setting is clothing recognition and manipulation. We propose a

method that enables a general purpose robot to bring clothing articles into a

desired configuration from an unknown initial configuration. Our method uses a

library of simple clothing models and requires limited perceptual capabilities.

Next, we present BigBIRD (Big Berkeley Instance Recognition Dataset), which has

been used in several areas relevant to cloud robotics, including instance

recognition, grasping and manipulation, and 3D model reconstruction. BigBIRD

provides 600 3D point clouds and 600 high-resolution (12 MP) images covering

all views of each object, along with generated meshes for ease of use. We also

explain the details of our calibration procedure and data collection system,

which collects all required data for a single object in under five minutes with

minimal human effort.

We then discuss the Yale-CMU-Berkeley (YCB) Object and Model Set, which is

specifically designed for benchmarking in manipulation research. For a set of

everyday objects, the dataset provides the same data as BigBIRD, an additional

set of high-quality models, and formats for use with common robotics software

packages. Researchers can also obtain a physical set of the objects, enabling

both simulation-based and robotic experiments.

Lastly, we discuss Brass, a preliminary framework for providing robotics and

automation algorithms as easy-to-use cloud services. Brass can yield several

benefits to algorithm developers and end-users, including automatic

resource provisioning and load balancing, benchmarking, and collective robot

learning.