IEEE Robotics & Automation Magazine - September 2018 - 35

performance suggests that ATRIAS can walk stably without
significant sensitivity to surface dynamics.
For our demonstration, we aimed to show locomotion
on rough ground without any vision or prior planning. To
create uneven ground in the laboratory, we tested walking
on various arrangements of stacks of plywood. Figure 10(d)
[50] shows the robot walking quickly (1.8 m/s) on a ran-
domly structured obstacle (maximum height 9.5 cm), com-
ing to a controlled stop at the end of the structure. The most
extreme laboratory obstacle tested was a 15-cm-tall plat-
form. In 11 consecutive tests, the robot successfully stepped
onto this platform, walked a few elevated steps, and stepped
off [shown in Figure 10(f) and (g)], [52]. Because the robot
was unable to plan for the obstacle, some of the foot place-
ments were not clean, including one test where the robot
landed on the obstacle on the point of its toe. The control
algorithm was able to recover in spite of these unexpected
contact modes and timings. Furthermore, in an outdoor
test, the robot was able to walk up and down a 15° slope
[Figure 10(c)], [49].
We also tested ATRIAS's response to unexpected distur-
bances, such as repeated dodgeball strikes [Figure 10(h)],
[53]. To deliver a much larger test impulse to a human-sized
robot, we gave the torso a series of firm kicks [Figure 10(i)],
[54]. When stepping in place, the robot was able to recover
from kicks imparting 60 kg·m/s of momentum without fall-
ing. (The size of the impulse delivered was inferred from sim-
ulating impulse disturbances in the high-fidelity simulator.)
This impulse is the equivalent of instantaneously accelerating
the robot to 1 m/s.
Speed
ATRIAS was able to match commanded speeds between zero
and 2.5 m/s and performed similarly well in both the for-
ward and reverse directions, though we noted that the robot
had the ability to achieve slightly higher speeds in the left-
ward direction as depicted in Figure 10(j) [55]. The latter fig-
ure shows a photo of ATRIAS reaching its top speed of
2.5 m/s (9 km/h) in an outdoor test on an asphalt path. After
accelerating faster than 2.0 m/s, short aerial periods with no
ground contact emerged, resulting in a transition to a run-
ning gait. This ability to transition between walking and run-
ning gaits was accomplished without switching between
controller structures. Figure 10(k) shows a snapshot of ATRIAS
after a transition to running during an outdoor test on artifi-
cial field turf, and Figure 10(l) illustrates the corresponding
ground-reaction forces measuring the length of the aerial
phases (an average flight time of 30 ms). This test also dem-
onstrated the robot's ability to accelerate from rest to a run,
and then to execute a controlled stop.
Energy Economy
We measured ATRIAS's energetic properties using two met-
rics: its operation time on a single battery charge and the
mechanical and TCoT [15]. To test battery life, we command-
ed ATRIAS to step repeatedly in place until the battery pack

was drained. The 48-V 10-Ah battery pack was drained in
approximately 30 min of operation.
The TCoT is a nondimensional measure of the energy
required to move a unit distance. The mechanical costs of
transport (MCoT) accounts for only the mechanical energy
being delivered by the actuators. The TCoT includes not just
the mechanical cost to locomote, but the resistive losses in the
electric motors and the onboard electronics overhead (in-
cluding wireless communication and the control computer).
We calculated the TCoT and MCoT for a 1.6-m/s walking
test of ATRIAS. On average, the TCoT was 1.3, as mea-
sured at the battery pack (current and voltage). This is an
improvement compared to the humanoid ASIMO's esti-
mated TCoT of 3.2 [22] but is still far from the TCoT of 0.19
reported for the Cornell Ranger [16]. The average MCoT is
0.96, as measured at the actuator outputs (torque and speed).
Discussion of Controller Behaviors
The broad effects of the three controller components (torso
balance, stride trajectory, and energy injection) can be seen
rather intuitively in the resulting behavior of the robot. In
the instance of directed perturbations, such as kicks and
dodgeballs [Figure 10(h) and (i)], the stride trajectory con-
trol was the most visible. The sudden velocity change from
the kick produced a significant velocity error, effectively
commanding a large recovery step. The imposed time limi-
tations between touchdown events in the stride generation
ensured that a new foothold would be taken before the
robot tipped too far. The effect of the torso balance control
was somewhat less overt to the naked eye but was also most
clear during the kicking experiment. After the initial per-
turbation and near the transition between the recovery
steps, the torso began to tip. However, once a new foothold
was secured, the torso snapped back to its vertical position
quite quickly.
The energy injection behaviors were most pronounced in
large [Figure 10(f) and (g)] or sustained terrain changes [such
as the hill climbing in Figure 10(c)] and when achieving fast
running speeds [Figure 10(j) and (k)]. After stepping on a tall
obstacle, significant velocity was lost, which registered as a
need to inject more energy and push up onto the obstacle.
When walking uphill, this additional push-off persisted and
added the gravitational potential energy necessary to continue
upward (and the reverse was true when descending). Fur-
thermore, when commanded to move sufficiently fast, enough
energy was injected through push-off so that the robot left the
ground. In essence, running manifested not as a distinctly pro-
gramed behavior, but as a necessary consequence of injecting
enough energy to locomote faster.
We also emphasize that the blending of the controllers was
critical to achieving the reported results. The smoothed tran-
sitions from stance to nonstance control were likely helpful in
randomly uneven [Figure 10(d) and (e)] or soft terrain [Fig-
ure 10(a), (b), and (k)]. By using force as a smooth criterion
for switching, the switch to stance was dependent on a firm
foothold being achieved. This meant that torso-balancing
september 2018

*

IEEE ROBOTICS & AUTOMATION MAGAZINE

*

35
IEEE Robotics & Automation Magazine - September 2018

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