robot and a real person, they become hesitant
to interact with it. “I try to create robots with a
humanoid shape but that are clearly not
human, because we are worried about that
uncanny valley,” says Kuffner.
When looking at the things we can get
a humanoid to do, it would be quite
difficult, for example, to leap from a singularly
tasked industrial robot to a humanoid that
could perform all the functions necessary to
build an automobile from the ground up
given the tools and parts, Kuffner notes. A
sufficiently skilled mechanic (or two) might be
able to do this.
However, because robots are computers
and computers are great at searching against
a database, it is highly conceivable that
humanoids can quickly become skilled at tasks
where optimal choices can be discerned from a
list in a database.
Dr Kuffner is optimistic about the
advancement of robots using human tools.
“The technology is improving a lot in the last
decade or so in terms of perception and sensing.
In terms of high resolution cameras, laser
scanners are now coming on line. Outdoor robots
didn’t work at all until we had reliable GPS,”
As these perception capabilities and
computing powers improve, using human tools —
having a humanoid pick up a cleaning instrument
such as a towel and wipe down a surface — could
happen in 10 years, Kuffner believes. There are
already prototypes that do that.
This 3D image shows ASIMO scanning the floor in front of him, searching
and discovering the best footsteps to achieve the desired goal.
foot. It processes tens of thousands of possible
footstep locations per second. The footstep
location that is the most highly safe and highly
effective toward reaching the ultimate goal
would be selected.
The algorithm also uses heuristics and
machine learning so that every time the robot
learns a bad place to step, it won’t step into a
similar place in the future, according to Kuffner.
The same algorithm has since been applied to
Kuffner’s Work on ASIMO
A 3D drawing of a humanoid using a footstep planning algorithm
and images from a right and left camera to plan a sequence
of steps toward the goal location.
Kuffner has been working with ASIMO
to achieve efficient walking. To accomplish
this, he has been developing a footstep
planning algorithm that searches possible
footstep choices over and over for the best
choice to avoid slipping, tripping, and bumping
into objects while walking a route to a desired
The algorithm looks at the state of the
floor and searches all the possible actions [steps],
then selects a sequence of actions to take to
reach the goal, Kuffner explains. “Imagine you
are on rough terrain and you are humanoid and
you want to climb to the top of a hill,” Kuffner
The humanoid would need to plan a
sequence of footstep locations to walk on to
reach that goal. The footstep planning algorithm
looks at the reachable region of ground for each
SERVO 02.2010 11