IEEE Robotics & Automation Magazine - September 2015 - 45

later provide better models for these objects using algorithms
that take advantage of the high-resolution RGB images for
building models.
In total, for each object, we provide the following:
● 600 RGB-D images
● 600 high-resolution RGB images
● segmentation masks for each image
● calibration information for each image
● texture-mapped 3-D mesh models.
The object scans can be found online at [55].
Models
Based on the scans of the objects, there are several ways in
which object models can be easily integrated into a variety of
robot simulation packages. For example, in the MoveIt [5]
simulation package, the mesh can be used as a collision object directly. Furthermore, a unified robot description format
(URDF) file can be automatically constructed to integrate
with ROS [56]. This provides a way to specify mass properties and can link to alternate representations of the mesh for
visualization and collision. Integration with the OpenRAVE
[57] simulation package is similarly straightforward where
we link to the display and collision meshes from a KinBody
XML file. Using the scans, we have created URDF and KinBody files for all of the objects in the data set, provided
alongside the scans at [55].
Once in a simulation environment, a variety of motion
planners and optimizers can use these models either as collision or manipulation objects. Some algorithms, such as Covariant Hamiltonian Optimization for Motion Planning [58],
require signed-distance fields to avoid collisions, which can
be computed from the included watertight meshes. Other
cases, such as Constrained Bi-directional Rapidly-Exploring
Random Tree [59], compute collisions directly using an optimized mesh collision checker.
In many cases, collision checking is a computational bottleneck for motion planning. Execution time can be reduced
using a simplified mesh produced either by hand or with automatic decimation methods [60]. We have not yet provided

simplified meshes in this data set, but we view this as an opportunity in future work to further explore mesh approximation algorithms and their impact on motion-planning
problems using the standardized benchmarks.
Functional Demonstration of Integration
into Simulation Software
The entire pipeline is shown in Figure 11. Here, we see the
HERB robot [61] preparing to grasp the virtual drill object. This
demonstration uses an integration of ROS and
OpenRAVE. The ROS is
A variety of motion
used to provide communication between the various
planners and optimizers
hardware and software
components of the robot,
can use these models
while OpenRave handles
planning and collision
either as collision or
checking.
Inside OpenRAVE, the
manipulation objects.
HERB robot uses CBiRRT,
the Open Motion Planning Library [62] library, and CHOMP to plan and optimize
motion trajectories. Using these tools, chains of several actions
can be executed in sequence. The simulation environment also

(a)

(b)

Figure 10. The point-cloud and textural-data overlays on two
YCB objects: (a) the mustard bottle and (b) the power drill.

(a)

Figure 9. The BigBIRD object-scanning rig: the box contains a
computer-controlled turntable.

(b)

Figure 11. (a) The screen-capture from the OpenRAVE
simulation and planning environment showing the HERB robot
[34] planning a grasp of the power drill object in the set. (b) The
actual grasp being executed by the robot on the physical object.

September 2015

IEEE ROBOTICS & AUTOMATION MAGAZINE

Table of Contents for the Digital Edition of IEEE Robotics & Automation Magazine - September 2015