IEEE Robotics & Automation Magazine - December 2016 - 101

physically interact with each other to identify which objects
supported other objects; second, we would use the created
models to make an optimal decision regarding which object
would be the safest to remove.
The high-level unloading sequence planner was
incrementally developed by progressively relaxing the
assumptions on the created world models. In [12], we
considered solid cuboid-shape objects (i.e., carton
boxes) with known poses and models, and we proposed
a method based on geometrical reasoning and static
equilibrium analysis to extract the gravitational support
relations between objects and build a relational representation that could be further used with a high-level AI
reasoning module to select the safest object. In [13], we
assumed that only the shape and pose of a subset of
objects composing the environment were available, and
we proposed a probabilistic decision-making framework
based on a possible worlds representation principle and
a machine-learning approach for probabilistic estimation of the gravitational support relations. In [14], we
presented an extension of our method to address the
problem of autonomously selecting the safest object
from a pile, considering objects with rigid convex
polyhedron shapes and examining the performance of
different machine-learning models for estimating the
support-relation probabilities.
Motion Planning
For facilitated motion planning, collision avoidance, and
manipulator control, we used a customized version of the
MoveIt! motion-planning framework. One of the main
reasons for modifying the MoveIt! framework is that it was
originally developed under the assumption that successfully planned trajectories are directly executed by the robot.
This assumption does not hold in our application scenarios, as we needed to plan both approach and escape trajectories prior to execution, and we might have needed to
attempt a multitude of target-grasp poses per object. These
requirements, in turn, necessitate a consistent handling of
the planning-scene objects and a careful managing of the
process of attaching and detaching object-collision geometries to the kinematic chain. We handled this problem by
employing a preplanning process, including a set of alternative approach/retrieve trajectories planned under different start/end pose constraints.
Grasp Planning and Execution
A reliable grasp acquisition of heterogeneous goods from
unstructured scenes is a challenging problem. The central
idea of our approach, as presented in [15] and [16], is to
exploit the low pregrasp pose sensitivity and the active surfaces of the velvet fingers gripper in the grasping process.
The experiments reported in [15] showed that, in cluttered
scenes, fingertip grasps are more likely to be feasible than
robust enveloping grasps, because the latter necessitate
large opening angles resulting in bulky gripper silhouettes

Figure 6. A sequence of intermediate grasp states in which the belts
of the gripper are used to pull the object toward its palm, which
results in a transition from a fingertip to an enveloping grasp.

for which no collision-free approach trajectories can be
found. We employ a simple pull-in strategy (Figure 6) that
exploits the underactuated nature and the conveyor belts
on the grasping device to embrace the object in a firm
envelope grasp by simultaneously squeezing it while actuating the belts inwards. The corresponding grasping controller was implemented by means of low-level current
control of the gripper's actuators, allowing for simple compliant behavior. The grasping strategy was implemented in
three steps: 1) the fingers were closed with a low current
set point until contact was detected; 2) the belts were actuated inward, while the gripper-closing DoF was kept in a
compliant low-current control mode, allowing the object
to be pulled into the gripper; and 3) if the phalanges had
wrapped around the object, a higher current set point was
commanded to ensure a firm grasp.
To autonomously achieve a grasp on an object, the
grasp-planning problem (i.e., finding an appropriate grasp
configuration and corresponding joint trajectories) needs to be
solved. We employed a data-driven solution where, to handle the
curse of dimensionality, the
grasp-synthesis problem
(i.e., finding a suitable palm
For facilitated motion
pose and gripper joint
configuration) is separatplanning, collision avoidance,
ed from the problem of
planning collision-free
and manipulator control, we
motions for the grippermanipulator chain. In an
used a customized version of
offline stage, the database
of known objects is poputhe MoveIt! motion-planning
lated with a set of fingertip
grasps that is synthesized
framework.
by following an approach
similar to [17], i.e., by
minimizing an energy function depending on the distance
and the alignment of the object relative to predefined
desired contact locations on the gripper's fingers. Additionally, the grasping principles observed in humans (i.e.,
the approach along an object-surface normal and orientation of the hand's lateral-axis normal to one of the object's
principal components) are incorporated by imposing
appropriate constraints to the underlying optimization
problem. The approach was implemented in the GraspIt!
(http://www.cs.columbia.edu/~cmatei/graspit) framework
and subsequently used to plan a set of 400 fingertip grasps
for each object in the database. As we could not employ
this procedure for objects that are not in the database, we
employed a two-step online approach to synthesize grasps
December 2016

*

IEEE ROBOTICS & AUTOMATION MAGAZINE

*

101
http://www.cs.columbia.edu/~cmatei/graspit
Table of Contents for the Digital Edition of IEEE Robotics & Automation Magazine - December 2016
