IEEE Robotics & Automation Magazine - June 2015 - 47

same space where they have been constructed. They have also
not been tested in manipulation challenges. In the context of
dynamic and flexible manufacturing floors, however, changes
may occur in the environment between the times that the
roadmap was constructed on the cloud and when it is queried
by the robot. Thus, it is interesting to evaluate the extent to
which the benefits of compact motion-planning representations apply to the case of a changing environment.

The pick-and-place subtask for each object is solved by
querying a multimodal manipulation roadmap, computed as
described in the "Planning Using Roadmap Precomputation"
section in three different ways (PRM), IRS, and SPARS), which
produces graphs of different density and size. Each pick-andplace manipulation problem is split into three subproblems: 1)
a transit plan from an initial configuration of the
arm to a grasping configuNeither the shelf nor the
ration of the movable
object on the horizontal
objects will always be
surface, 2) a transfer plan
between the reachable
accurately placed, and,
grasping configuration of
the initial object location
locally, the robot needs to
to a grasping configuration at the goal, and 3) a
adapt to the current state
transit plan from the
grasping configuration at
of the scene.
the goal to the safe initial
arm configuration.
The higher-level task planner attempts three different randomly sampled grasps at the initial and goal object positions
and evaluates all nine combinations to find the best solution
given the available roadmap. This means that, for a single
pick-and-place task, the transit roadmap is queried six times
(three ways to reach the object at the initial location and three
ways for the arm to be retracted from the goal placement).
The transfer roadmap is queried nine times, i.e., for all combinations between grasps in the initial and goal placement of
the object. The reported online query resolution time is the
time taken to transfer all three objects, which corresponds to
18 queries of the transit roadmap and 27 queries of the transfer roadmap.
The memory footprint of the roadmap increases with the
number of iterations and is the aggregate of all vertices and
edges in the roadmap. The PRM) roadmap is the biggest and

Evaluation of Methods and Tradeoffs
This section evaluates PRM), IRS, and SPARS in terms of computational efficiency, path quality, and success ratio for a changing scene. The setup shown in Figure 5 involves a seven-degreesof-freedom Baxter arm that must transfer three rigid bodies.
The experiments were performed with the help of simulation
software [12]. In a situation similar to the one in Kiva Systems'
warehouses, mobile robots may bring a shelf containing products to an area where they will be packed. While today this task
is performed by people, it is envisioned that, in the future, robot
manipulators will be able to address this challenge. Such robots
will have access to significant computing power and precomputation can be employed to improve their performance. Neither
the shelf nor the objects will always be accurately placed, and,
locally, the robot needs to adapt to the current state of the scene.
Inspired by this scenario, there are three variations of the
problem considered in the experiments performed here: an easy,
a medium, and a hard challenge. The easy challenge involves
objects placed on a tabletop surface. This is the environment
that is provided to the roadmap during the precomputation
phase. The medium-difficulty problem has the objects placed on
a shelf that is not available during roadmap construction. The
hard instance has the objects placed far back on multiple shelves,
close to obstacle geometry in a way that complicates their grasp.
The objects are initially placed on a horizontal surface in a vertical configuration and must be placed in a horizontal configuration inside a box. For a challenge to be considered solved, the
robot must manage to transfer all three objects.

(a)

(b)

(c)

Figure 5. The three versions of the environment used for evaluating the methods, inspired by the Amazon Picking Challenge: (a) easy,
(b) medium, and (c) hard. The offline roadmaps are computed given the easy environment. This problem involves a Baxter arm grasping
an object from the tabletop in a vertical configuration and placing it inside a box in a horizontal configuration. The medium challenge has
the objects placed on a shelf that was not known during the construction of the roadmap and limits the motion of the robot. The hard
challenge has multiple shelves, with the objects placed further back than in the medium challenge, which further limits the capability of the
arm to grasp the items.

June 2015

IEEE ROBOTICS & AUTOMATION MAGAZINE

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