Whole Body Control
The biggest engineering challenge has
been bringing all the disparate teams
together. "A small change on what
the software team needs in terms of
computing power can ripple through
the entire mechanical design," says
Another challenge has been using
parts in ways that have never been
done before. For example, the main
motor is a series elastic actuator, the
same type of motors that are used in
modern prosthetics, which have never
really been applied to a full sized
The robot's actuators were also custom made. According to Seminatore,
what makes them clever is that they
are not rigidly attached to the robot;
instead, they are attached via titanium springs for shock absorption.
"It functions like your tendons," he
adds. "Your muscles aren't directly
connected to the bone, you've got the
springiness of the tendon that helps
store energy and make walking much
Traditionally, robots are controlled
via position control. You command
a joint to go to a certain angle and
then you play that over and over. In
ESCHER, load cells are incorporated
into each of the motors which measure the force placed on each of the
joints. This gives the robot a basic
sense of touch as it walks. It reaches
its foot out, slowly puts weight on it,
making sure there is a stable plant,
then moves the next foot.
Additionally, the robot features a custom whole-body controller which gives
the robot center of mass awareness.
"If I shove the robot and that center
of mass moves, the robot can calculate how much force it takes to move
its center of mass back into position
because we can calculate all the forces going through each of the joints,"
If the robot were to be shoved from
behind, the robot moves forward, drops
its hips, flares its arms back and then
stands up, much like human nature.
Developing that center of mass was a
difficult task. In fact, the TREC lab may
be the second or third group in the
world to get this capability working on
a robot. "We were right there in the running with those who did it for the first
time," he adds. "Normally you focus on
controlling a specific joint, we're focusing on what the whole body is doing
when a joint moves"
A SAFFiR Robot
All of the work the team has done in
preparation for the DRC will help TREC in
their SAFFiR project for the Navy, as the
ESCHER platform will eventually replace
the current SAFFiR robot. The next challenge, apart from the DRC, will be to
improve mobility on a ship.
"One of the big problems a lot of the
DARPA robots have is that they are big,"
explains Seminatore. "We are on the
smaller side compared to a lot of the
other robots, and that's by design." This
design is necessary as the SAFFiR robot
will have to fit into much smaller, much
more confined spaces of a ship, and handle a moving floor with no outside frame
of reference. The robot will also have to
work around people.
"The firefighter or sailor that
is operating near the robot
has to be able to shove the
robot out of the way to get
past it and not be afraid to be
in a very confined space with
it," says Seminatore. An all-electric robot, ESCHER is safer
to operate around than its
However, there are still a
huge amount of challenges
to overcome before the robot
is "grunt proof." Right now, it
takes a team of PHDs to oper-
ate ESCHER. It still needs to
be simplified down to something more
robust, durable, and easier to interact
"You don't want a massive computing
station and a command center operating a single robot," says Seminatore.
"Even a joystick is probably a little too
much," he adds.
Other challenges pertain to actions
generally taken for granted. For
example, picking up a cup from a table.
"When I reach out and I try to pick up
the cup, I know if there is water in the
cup. I shouldn't turn it upside down,"
says Seminatore. "I know the best way
to approach the cup." The robot has no
idea; it doesn't even know that the cup
isn't attached to the table.
Today, the best algorithms can only
identify objects with a 60% accuracy
using supercomputers, but a robot can’t
carry that much computing power on
In addition to knowing the object, the
robot must also know its condition.
"How does the robot know that the cup
can't be picked up from the top because
it needs access to what's inside of it,"
These are enormous artificial intelligence challenges, and while the TREC
lab has been working to overcome
them, they still have a long way to go.
"It’s easy to design a robot that people
find very difficult," says Seminatore. "I
can design a robot that can lift 2,000
pounds no problem ... but it’s the simple
things that robots really have a lot of
The lab demonstrated SAFFiR to the
Navy in November and is ramping up for
Phase Two. Seminatore doesn't expect a
finalized product for ten to 15 years, but
in the mean time, they will take to the
robotics challenge with ESCHER hoping
to put on a good show, demonstrating
what the future may hold for the Navy’s