IEEE Robotics & Automation Magazine - June 2015 - 44

the environment. The computation of the manipulation path is
computed by a motion-planning algorithm that resides exclusively in the cloud. Upon the computation of a path, the cloud
transmits this solution to the robot, which follows the cloudcomputed path. Thus, the
communication loop in
this scenario involves the
The question is how the
robot transmitting sensing
data and receiving solution
local workstation can best
trajectories.
This alternative is a
interact with the largerpotentially viable way of
incorporating cloud comscale computing cloud so
puting into robotic operathat the manipulation task tions but also raises some
concerns. It requires the
is accomplished efficiently. frequent transmission of
large amounts of data to
the cloud, including the
frequent communication of 3-D point clouds. Furthermore, in
this scenario, the robot is fully dependent on the cloud and the
reliability of this communication. Given a communication failure, which can occur even in a relatively reliable industrial
environment, or any other issues that may arise in interacting
with the cloud, the robot is not able to plan and react to the
changes in its environment.
Synergistic Robot-Cloud Computation
The aforementioned issues with cloud-only computation
illustrate the need for a hybrid solution, where some computation takes place on the local workstation but the cloud is
responsible for the more computationally heavy but not
time-critical operations. Such a solution should result in 1) a
reduced communication load relative to cloud-only computation and 2) increased responsiveness to environmental
changes relative to the local-only computation, where the

Transfer
State Space

Transit
State Space

Transition States
States Where (Stable Grasps) States Where the
Object Is in
the Manipulator Is
a Stable
Grasping an Object
Configuration
Configuration

Figure 3. The constructed roadmap needs to span the space of
manipulation planning.

IEEE ROBOTICS & AUTOMATION MAGAZINE

June 2015

robot is limited by the computational capabilities of an individual workstation.
In the synergistic mode of operation, the robot only
occasionally transmits sensing data to the cloud at a low
rate. The computing cloud must then identify the static
obstacles and preprocess the collision-free configuration
space of the robot. The result is a motion-planning data
structure, i.e., a roadmap, which contains collision-free
paths based on the static geometry and can be used for
manipulation purposes. When computed, the roadmap is
transmitted from the cloud to the robot's workstation,
where it is used to quickly answer queries.
The local workstation has access to the most updated
sensing information, i.e., the latest 3-D point cloud, which is
more recent than the data available for the roadmap's construction on the cloud. The local workstation does not just
blindly follow the shortest path on the roadmap. Instead, it
finds the shortest path that does not result in collisions with
the latest available sensing data. Thus, the language of communication between the robot and the cloud becomes the
motion-planning data structure, i.e., the roadmap, which is
updated periodically on the cloud using the sensing updates
and transmitted back to the robot at a low frame rate.
An effective implementation of the synergistic model
should aim to achieve the following.
● Computational efficiency: It should minimize the time to
compute solutions for motion-planning queries on the local
workstation relative to local-only computation. This
requires taking advantage of the cloud-based preprocessing.
● Communication overhead: It must reduce the communication overhead of the robot-cloud interaction relative to the
cloud-only computation. This should help to reduce the
sensitivity of the approach to bandwidth, latency, and
throughput challenges.
● Safety and reactivity to changes: It should return safe solutions, where the robot is able to safely avoid any obstacles
that are detected by its sensors but were not known to the
cloud when the roadmap was constructed.
● Path quality: It should compute high-quality solutions,
where the manipulator moves in an efficient manner without executing redundant motions that will delay the execution of the task.
Planning Using Roadmap Precomputation
The following discussion focuses on motion-planning primitives from the related literature that can be used to achieve a
proper balance among the four previously stated objectives in
the context of a synergistic robot-cloud computation.
Computing a Manipulation Roadmap
The planner executed on the cloud can explore the task's state
space by generating a manipulation roadmap R^V, Eh, given
the robot and an object that needs to be manipulated [21]. The
geometry of the object is assumed to be known, but the configurations in which it will appear are not. The method operates over collision-free states of the form x = ^^q, p h, t h, where

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