IEEE Robotics & Automation Magazine - June 2013 - 41

system is called "computed unified device architecture
(CUDA)."
The CUDA approach has gained much attention, and
many researchers have found it a powerful platform for
boosting their computations. Furthermore, several libraries,
such as CUBlas (a port of the Basic Linear Algebra Set-Blas)
and GpuCV (largely compatible with OpenCV) for computer
vision, have allowed researchers to take advantage of the computing power of GPUs without requiring an explicit parallelization of algorithms. Applications such as MATLAB or GNU
Image Manipulation Program (GIMP) are also provided with
CUDA extensions that permit the applications to transparently benefit through GPU processing.
We believe that the robotics community should also benefit by adopting and using such technology. To this aim, three
main lines are followed:
● parallelize and port existing algorithms to execute in GPUs
● take advantage of already developed general-purpose math
or computer vision libraries and tools
● develop new algorithms explicitly taking into account the
computational capabilities of such devices.
Many algorithms in mobile robotics are computationally
intensive. Among them, the map-building or simultaneous
localization and mapping (SLAM) problem [1], [2] has
gained great attention over the last few decades with a prototypical case study of indoor wheeled mobile robots equipped
with laser rangefinders.
The following section presents current related work on
GPU robotic applications; note that most contributions are
related to the computer vision domain. This article presents a
demonstration of GPU computing applied to the SLAM
problem as the three lines described previously, and also
shows its high potential applicability in domains other than
vision. The section describing grid mapping shows how a
publicly available implementation of the grid-mapping algorithm can be parallelized over a GPU to obtain high computational savings. The section on graph optimization uses a
general GPU-optimized sparse linear system solver to address
the graph SLAM problem, defined as a least squares minimization over a graph of poses. The section discussing grid
matching implements a novel algorithm for handling overlapping between different grid maps, which can be efficiently
applied because of the computing power of the GPU. The section describing experiments reports on simulated and real
experiments combining some of the techniques previously
mentioned. A discussion on related issues is presented, and
conclusions are summarized.
GPU Computing in Robotics

vision domain, as in the early work of Michel et al. [4], which
tracked 3-D objects with cameras using GPUs to achieve realtime performance when controlling a humanoid. Other
works, such as [5], try to speed up the typical processing and
matching of scale-invariant feature transform features among
different frames for localization purposes, whereas more
complete multimodal perception approaches, such as [6],
include Bayesian solutions with particle filters. There exist
research groups and projects fully devoted to this area, such as
the Gpu4Vision Project [7].
In the domain of mapping with laser rangefindThe CUDA approach has
ers, the work of Yguel et
al. in 2007 [8] addressed
gained much attention,
the problem of updating a
2-D probability grid with
and many researchers
a novel formulation of the
required polar-to-Cartehave found it a powerful
sian grid conversion that
takes into account the
platform for boosting
actual beam model. More
recent work is found in
their computations.
the well-known Slam6D
open-source project [9],
where NVIDIA CUDA is used for speeding up 3-D point
cloud registration and iterative closest point matching [10].
It is important to highlight the merits of [4] and [8].
Because the current general-purpose CUDA tools were
not available earlier, programmers had to deal with
specific graphics application programming interfaces.
Today, these tools allow much more simple development,
and even robotic-specific software frameworks include
support for such tools like, for example, ROS [11] does
with CUDA.
Overview of NVIDIA CUDA Architecture
CUDA exposes the NVIDIA multicore GPUs computing
capabilities through the following elements (Figure 1):
● Thread hierarchy-The execution unit in CUDA is a
kernel, which is structured in a so-called grid (a onedimensional or 2-D array) of blocks, each block in turn
arranged in another (up to 3-D) array of threads. Unlike
CPU multithreading, every thread of the same kernel has
to run exactly the same code, so typically a kernel is used

Block

Related Work
In recent years, there has been an upsurge of interest in GPU
computing applied to robotics. Some applications address
topics such as grasping with manipulators and solving the
algebraic and geometric problem with GPUs [3]. Most of the
related literature, however, is actually found in the computer

Fast
(Registers)

Kernel Grid Threads

Fast (Read Only)
Fast (Cache-Like)

Block

Slow

Device Memory
Local (Thread)
Constant
Shared (Block)
Global

Figure 1. CUDA architecture.

june 2013

Ieee ROBOTICS & AuTOMATIOn MAGAZIne

Table of Contents for the Digital Edition of IEEE Robotics & Automation Magazine - June 2013