IEEE Robotics & Automation Magazine - June 2012 - 66

advance and also stored in the database. For each object in
the database, we used the GraspIt! simulator [28] to precompute a large number of grasp points for the PR2 gripper. Using the grasp quality function and search method
described in [27], we performed offline grasp planning for
the objects in our database, allowing 4 h of planning time
per object. This process resulted in an average of 600 precomputed grasp points saved in the database for each
object, as exemplified in Figure 10.
We use the database of known objects and grasps in
conjunction with an object recognition module that
attempts to match each object point cluster, segmented as
described in the previous section, to each of the object
meshes in a predefined set. Matching is performed using
an iterative technique similar to the iterative closest point
algorithm [29], and the best fit is returned. Other recent
recognition methods, such as the object recognition and
pose estimation for manipulation framework [30], have
shown the ability to recognize objects even when stacked
closely together, and could be used to allow the robot to
perform grasps in highly cluttered scenes.
Grasp Execution
The grasps generated by either of the two planners discussed
above consist of end-effector poses relative to the target.
After planning, a collision-aware IK module checks each
grasp for feasibility in the current collision environment,
constructed as described in the previous section. Once a
grasp is deemed feasible, the motion planner generates an
arm trajectory for achieving the grasp position (as described
in the "Environment Modeling, Motion Planning, and Execution" section), and the grasp is executed.

(a)

(b)

Coping with Uncertainty
The methods described so far can be used to grasp a wide
range of common household objects, not only taking
advantage of high-level information (such as object identity) when it is available but also handling novel objects
when needed. In practice, however, we also found them to
be sensitive to errors in either perception or execution
modules, where an object was misrecognized, segmentation was incorrect, or calibration between the sensors and
the end-effector was imperfect.
Probabilistic Grasp Planning
One method of dealing with uncertainty in object shape or
pose due to such errors is to combine information from
multiple object recognizers, grasp planners, and/or grasp
evaluators. Figure 11 illustrates the general concept:
when trying to grasp the cup shown in Figure 11(a), the
robot's stereo cameras see only the point cloud shown in
Figure 11(b). The robot's object recognizers provide two
hypotheses, with the corresponding quality scores: the tennis ball can (incorrect) shown in Figure 11(d) and the cup
(correct) shown in Figure 11(e). Grasps can be planned on
all possible hypotheses, including different object shapes
and poses (using, for instance, GraspIt!) and even just the
segmented point cluster (using, for instance, the cluster
grasp planner described in the "Grasp Planning and Execution" section).
The entire pool of grasps generated on all available
hypotheses can be evaluated using any number of available
grasp evaluators. In this case, the grasp evaluators used are
the same ones used to generate the grasps: GraspIt! It can
be used to evaluate arbitrary grasps on each possible object

(c)

Figure 10. Grasp planning in a simulated environment for a known object: (a) the mesh, (b) a simulated grasp, and (c) the
complete set of precomputed grasps.

IEEE ROBOTICS & AUTOMATION MAGAZINE

JUNE 2012

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