IEEE Robotics & Automation Magazine - September 2018 - 37

parameters that worked for the full-order and heavily non-
linear system.
Exploit Rapid Iteration
The key to developing a working product, both hardware and
software, is a quick design-test-evaluate cycle. Incremental
testing allowed us to quickly go over ideas, discover what
worked, and cut fruitless branches out of our search.
Allow for Adjustable Control Parameters
Shifting model parameters are a fact of life in robotics. Not
only does ATRIAS wear and age like any robot, it can also
behave slightly differently with every test. Trimming controls
like those for a radio-controlled airplane helped balance the
robot on different terrain, with slightly different link lengths
due to manufacturing, and different amounts of onboard
weight. These tuning parameters were built into the controller
and could be adjusted on the fly.
Be Careful Not to Actively Control Any Behavior That Is
Actually a Symptom of a More Subtle Control Target
This difficulty occurs frequently in the field of bipedal loco-
motion, i.e., the most important feature of a gait appears to be
the center-of-mass (CoM) motion or ground-reaction forces,
so many robots try to control exactly those. For highly under-
actuated and dynamic robots, controlling around a trajectory
is extremely difficult; the control authority of the robot is lim-
ited and phase dependent. Simply by choosing a different
control target, e.g., stride length, angular momentum, or ver-
tical impulse, periodic CoM motions emerge naturally. Sim-
ple control targets generally require no preplanning and are
reactive to changes in the environment, but they still excite
the natural walking dynamics of the robot.
Moving Forward: Future Iterations
ATRIAS can walk and run at various speeds and on varied ter-
rain, but has incredible difficulty standing still. The robot has a
minuscule polygon of support, and active stabilization tests
suggested just a tiny region of stability, even in simulation.
(Investigations into using linear quadratic regulators to locally
stabilize a fixed point associated with standing yielded an
impractically small basin of attraction.) While ATRIAS can
hold its position by stepping in place, this is not an energetically
practical solution for idling. In future designs, having an ability
to apply even limited stabilizing torques about the foot, while
not impeding gait dynamics, would be helpful for stationary
balancing, climbing stairs, and precise balancing between steps.
Turning is a major component of locomotion, but ATRIAS
can only sidestep. Animals are able to zigzag and bank between
obstacles and points of interest, and we want future robots to
have that same ability. To do this, ATRIAS would need an
extra actuator to control the long-axis rotation of the leg,
applying yaw torques to the ground and turning the robot.
Currently, the robot requires the human operator to manually
steer via a carbon tube extending from its torso, and turning the
robot breaks static friction between the feet and the ground.

Practical robots will need to be self-starting and self-park-
ing. ATRIAS must be started from a hanging position, and
the shutdown process effectively stops the robot in midair
and causes it to fall. Future iterations should be able to stand
alone from a parked position and return to that position
when shut down.
Escalating emergency states could gracefully handle small
errors without a full E-stop. Currently, there is only one E-stop
case: shut down all motor drivers, crashing the robot. This case
is triggered for everything from overheating motors and limit-
switch triggers to a dangerous controller failure. We disabled
many of these safeties for the live shows at the DARPA Robot-
ics Challenge to prevent unnecessary halts.
Efficiency can continue being increased. ATRIAS uses
Harmonic Drive gearheads for their small package, but they
are extremely inefficient. In addition, an internal power loop,
where one motor acts as a brake, unnecessarily dissipates
energy. Leg design must analyze the task force and speed
requirements as they relate to the mechanism kinematics,
minimizing the work lost to self-braking [3], [25].
The real world is a chaotic and dangerous place for robots,
but machines with increased autonomy will need to be able to
withstand and recover from crashes. ATRIAS requires a safety
tether to prevent it from falling, because it was designed to be
a scientific demonstrator of spring-mass locomotion, not a
durable field-ready product. Future robots should be sturdy
enough to fall or to crash into trees.
Conclusions
The ATRIAS robot, with a combination of deliberately engi-
neered passive dynamics and complementary control algo-
rithms, was able to walk and run, without external power or
support, in front of a live audience at the DARPA Robotics
Challenge. In producing this live performance, ATRIAS dem-
onstrated 2.5-m/s running, variable speed control, and the
ability to recover from strong human kicks. Furthermore, the
robot was able to traverse varied surface dynamics and obsta-
cles as high as 15 cm without any planning or vision. To the
best of our knowledge, this degree of terrain robustness has
not been reported for a self-contained bipedal machine. What
ultimately allowed for sufficiently fast progress was a commit-
ment to rapid control iteration on hardware. However, by the
nature of its highly compliant and underactuated design,
every step along the way required ATRIAS to embrace its pas-
sive dynamics to keep moving forward.
Acknowledgments
This work was funded by the DARPA Maximum Mobility
and Manipulation Program, Grants W31P4Q-13-C-0099 and
W91CRB-11-1-0002; Human Frontier Science Program,
Grant RGY0062/2010; and the National Science Foundation,
Grant CMMI-1100232.
References
[1] M. H. Raibert, Legged Robots That Balance. Cambridge, MA: MIT
Press, 1986.

september 2018

*

IEEE ROBOTICS & AUTOMATION MAGAZINE

*

37
IEEE Robotics & Automation Magazine - September 2018

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