IEEE Robotics & Automation Magazine - June 2012 - 53

computation times. In the experiment, depicted in Figure
11, the quality of the best solution is determined over time
(see Figure 12).
Performance
The performance of the proposed Grasp-RRT planner
in single- and dual-arm planning setups is presented
in Figure 13 and Table 1. The run-time analysis has
been carried out on a dual-core CPU with 2.7 GHz by
averaging 50 test runs. The time spent for the three
main parts of the algorithm is distinguished, pointing
out that the parameter setup was well balanced since
approximately the same amount of time is spent for
building the RRT, computing the approach directions,
and evaluating the grasping poses. The last two columns of Table 1 show the number of approach trajectories that have been generated and the number of
grasp measurements that were calculated during the
planning process. These values differ since not all
approach trajectories result in a suitable grasping
configuration.

using precomputed reachability information similar to the
approaches presented in [8].
One of the advantages of the Grasp-RRT approach is
that neither a dedicated grasp planner nor a robot-specific
IK solver is needed. Furthermore, no precomputed reachability information is needed to speed up any IK queries.
In the 3-D mesh (Figure 14), the start and final configurations that were generated during planning are depicted.
In Table 2, the results of the RRT, IK-RRT, and Grasp-
RRT approach are shown. Initially, the three approaches
are compared in a scene where the target object is reachable without difficulty. In this situation, it was sufficient to
plan 50 grasps to achieve a collision-free and reachable IK
solution for the RRT and IK-RRT algorithms. The average
time for this step was measured with 2.7 s. Before planning
a collision-free motion with the RRT approach, an IK

0.20
Right Hand
Grasp Quality

0.16

Time (s)

Left Hand
Comparative Study
0.12
In the following experiment, we simulate the application of
0.08
the Grasp-RRT approach for grasping an unknown object
to compare the proposed approach with classical stepwise
0.04
algorithms for planning grasping motions [19]. Therefore,
we use an imperfect 3-D model of an object that was cre0.00
0
10
20
30
40
ated with the approaches of [24]. The object is located in
Time (s)
front of the robot, and the task is to create a collision-free
grasping motion without using any precomputed sets of
grasps. Hence, RRT- and IK-RRT-related planning Figure 12. The quality of the resulting grasp increases over
approaches must initially create such grasping information time. In this example, the first force-closure grasp was found
after 70 ms. When high-quality grasps are needed, several
by building a set of potential grasps. This set of grasps is seconds have to be spent until the quality of a resulting grasp
used before (RRT) or during (IK-RRT) planning to solve cannot be increased significantly.
IK queries for determining target
configurations. In this experiment,
grasp planning is performed by
15
computing a set of 50 feasible grasps
13.3
Grasp Scoring
with the wrench-space approach, as
Approach Movements
it is used in GraspIt! [10]. To allow
RRT Buildup
comparison with the Grasp-RRT
approach, the grasp planner that is
10
included in Simox is used so that the
7.6
7.2
performance measurement as well
as the resulting grasp quality is based
on the same source code. Further5
more, an efficient IK solver must be
4.0
present for the RRT and IK-RRT
planner to determine collision-free
0.6
0.6
IK solutions to one of the planned
grasps. In general, the majority of
0
Measuring Measuring
Wok
Wok
Bowl
Bike
the planned grasps are not reachable
Cup
Cup
(0.08)
(Single
Arm)
by the manipulator, we perform a
filter step to generate a subset of
reachable grasps. This is done by Figure 13. An overview of the average performance measurements.

JUNE 2012

IEEE ROBOTICS & AUTOMATION MAGAZINE

Table of Contents for the Digital Edition of IEEE Robotics & Automation Magazine - June 2012