IEEE Robotics & Automation Magazine - September 2018 - 32

the effect of the periodic forcing, but that impulse is being
continually injected, so it remains dominant. It is not neces-
sary for the controller to sense and act on current velocity
feedback; it must only supply feed-forward forcing in the
desired direction.

Our final controller occupied a concise 880 lines of
MATLAB code, not including the Simulink architecture. This
did not include microcontroller or low-level safety code, only
the final control function that mapped the robot's state to
motor commands.
Figure 9 shows the quality of the simulation, where the
robot was kicked during a physical test and given the same
impulse in simulation. The resulting robot trajectories were
nearly identical, from the overall torso path down to the
motion and timing of the leg movements.

Methods and Processes
Having such an atypical design and control paradigm for a
humanoid, ATRIAS required interactive tuning and gradual
addition of behaviors to add capabilities to the controller. A
high-fidelity SimMechanics model made it easy to continually
tune and adjust the controllers, with results that could be
immediately used on the robot.

Development of Behaviors and Capabilities
We started with an intuitive, bare-bones controller for two-
dimensional (2-D) push-assisted walking on a spherical
boom. Adding a rear leg push-off behavior allowed the robot
to walk on its own. This process continued, adding behaviors
to the simulation, tuning them, and applying them to the
robot. Soon, we had a controller that could stably walk and
run in 3-D through rough and unstable terrain. There were
several significant stages in developing this controller, start-
ing with a basic controller and incrementally adding hand-
coded behaviors.

Tools
Having a quality software toolchain was vital for reaching the
goal of a live show at the DARPA Robotics Challenge. A com-
bination of off-the-shelf software and hardware components
was selected and assembled to form a control system that
required minimal maintenance effort. As an added benefit,
the ATRIAS controller could be written as native MATLAB
code. This toolchain allowed for rapid iteration of controllers
and simple testing on the full-order robot.
A multibody simulation of the ATRIAS biped allowed for
extremely efficient and worry-free testing of new control
ideas [44]. Modeled in SimMechanics and controlled
through Simulink, the simulation was a good approximation
of the behavior of the real robot, down to the same controller
interface. A controller could be tuned using the simulation
and then required only minor adjustment when implement-
ed on the robot. The porting process was handled almost
entirely by the software, requiring only a flag indicating
whether the controller was running in the simulation or on
the hardware.

State-Based Push-to-Walk on Flat Ground: 2-D
The inaugural behavior was primitive stepping with a fixed
stride length, designed simply to put one foot in front of the
other. This behavior was not automatic, so an operator had
to be present to push the robot forward from step to step.
Stance leg and swing leg were determined by which physical
leg was applying greater force to the ground, a parameter
that typically switched when the hip was halfway between
footfalls. Forces in the leg length and leg angle were mea-
sured by sensing the deflection of the two series springs in
each leg.

60 kg . m/s

(a)

(b)
Figure 9. An example of the use of full-body simulation in the development of ATRIAS's control. (a) A film strip of the SimMechanics
model of ATRIAS with a hardware controller being simulated with a large horizontal impulse (aesthetically illustrated with an overlaid
kicking human). (b) Using the same controller, the physical robot is kicked to test its disturbance rejection. (Figure courtesy of
Christian Hubicki.)

IEEE ROBOTICS & AUTOMATION MAGAZINE

september 2018

IEEE Robotics & Automation Magazine - September 2018

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