IEEE Robotics & Automation Magazine - September 2018 - 33

To take the next step, the foot of the leading leg was
brought to a point above the next target and then lowered
until it contacted the ground. The foot trajectory from the
previous midstride to the next midstride was parameter-
ized by the x-location of the hip between midstrides. After
contact, the motors continued moving as if the foot was
still in free space, causing the springs to deflect and
apply a restorative force. Torso stabilization was achieved
through PD-controlled hip torques on the legs, scaled by
the vertical ground force of each foot (i.e., a scaled fric-
tion cone).
Self-Regulated Walking on Flat Ground
Incorporating a push-off behavior to the trailing leg allowed
the robot to propel itself forward, adding back lost energy.
Low proportional gains for the joint trajectories and high
derivative gains damped energy out of the gait and stabilized
the robot. The interaction between energy injection in the
forward direction and energy removal by the joints led to a
stable forward walking speed.
Stepping in Place: Incorporating a Clock-Driven
Stepping Cycle
The previous behaviors were not self-starting, so we added a
clock-driven stepping cycle. This meant the robot would
always be in motion, even at zero forward velocity. It began by
stepping in place, trying to maintain a fixed horizontal
position. Footfalls were selected to remove any extraneous
momentum by shifting the horizontal position of the step
proportionately with velocity error.

also made the strides fully dependent on the clock, removing
any state-based feedback.
Robust Stepping, Obstacles, Kicks, and Dodgeballs
Stability is the most important factor for real-world locomo-
tion, so we spent time making sure the stepping controller
could handle large velo-
city changes and altera-
Having a quality software
tions in ground height
and consistency. Our im-
toolchain was vital for
pulsive testing included
small pushes, a dodgeball
reaching the goal of a
barrage, and heavy kicks.
In separate tests, foam
live show at the DARPA
squares and wooden steps
disrupted the flat-ground
Robotics Challenge.
stepping cycle. The con-
troller's behavior took the
extraneous velocities from
these disturbances and removed them in several steps, set-
tling into a zero-velocity stepping pattern.
Directed Stepping
Just as stepping in place led to speed changes in 2-D, the
3-D stepping controller was given directional commands. A
video game controller influenced the velocity of the robot
by suggesting a direction of motion, which the robot tried
to satisfy. The speed change was not immediate, but the
velocity asymptotically approached the commanded speed
and direction.

Speed Changes, Forward, and Reverse
With the stepping behavior implemented, we began varying
the forward velocity command from positive to zero, to nega-
tive, and back. The robot began by stepping in place at zero
velocity and then slowly increased the forward speed to begin
walking forward. The stride length was varied, depending on
the forward speed. The push-off behavior was always in effect
as a result of the clock-driven stepping cycle.

Robustness with Obstacles
We performed another round of robustness testing for the
directional walking controller, this time in 3-D. We used
more unstructured terrain, with foam pads, blocks, and
plywood steps. At this point, we designed and built a light-
weight mobile gantry that could be pushed around by a
pair of researchers while a third drove the robot using the
game controller.

Numerous Obstacles and Stability Testing
At this point, the controller was able to robustly walk in mod-
erate step-ups, step-downs, loose terrain, and slippery terrain
in 2-D. It could also handle pushes and kicks that either accel-
erated or impeded its forward progress.

Running on Flat Ground
Given enough space, the controller could pick up sufficient
speed to enter an aerial phase. As there is no ground contact
in flight, we could not judge the forward position of the robot
through the stance leg. Instead, we gauged forward position
by integrating the last known forward velocity. Because for-
ward velocity in flight was constant and flight times were rela-
tively small, this approach worked well for maintaining
ground-speed-matched toes.

Stepping in Place: 3-D
The next big step in creating a controller for real-world loco-
motion was taking the robot off its support boom, which is
the walking equivalent of removing the training wheels from
a new cyclist's bike. Now, the robot had to control its lateral
velocity and torso roll in addition to forward velocity and
pitch. We modified the clock-based stepping controller to
account for these additional DoFs rather than adding entirely
new behaviors. The first test of this capability was simply
stepping in place with a zero-velocity goal. At this point, we

Locomotion Capabilities
We report ATRIAS's walking and running capabilities in
terms of robustness (as measured by both terrain variation
and external perturbations), speed, and energy economy.
These capabilities were assessed in a variety of experimental
tests in the lead-up to the DARPA Robotics Challenge.
september 2018

*

IEEE ROBOTICS & AUTOMATION MAGAZINE

*

33
IEEE Robotics & Automation Magazine - September 2018

Table of Contents for the Digital Edition of IEEE Robotics & Automation Magazine - September 2018
