IEEE Robotics & Automation Magazine - September 2018 - 28

absolute encoders at each internal DoF provide joint angles
and spring deflections, and a fiber-optic gyroscope provides
torso orientation. These sensors are sufficient to determine
the configuration of the robot, save for its translation with
respect to the world frame.
Custom, versatile interface modules, called Medulla mod-
ules, read, translate, packetize, and send proprioceptive and
orientation sensor data to the control computer. The Medullas
also read thermistors embedded in the motor assemblies so
the controller can detect and respond to overheating. Some
Medulla modules are used to pass torque commands to the
hip abduction motor drivers, as these cannot connect directly
to the EtherCAT bus.
The battery pack uses four six-cell lithium polymer battery
packs, each with a 5-Ah-rated capacity. The packs are con-
nected in a two-serial two-parallel configuration, giving a
nominal voltage of 44.4 V and a nominal capacity of 10 Ah.
With 65 C-discharge-rate batteries, the pack is rated to deliver
a peak current of 650 A.
A supervisory computer communicates with the robot
computer through a Wi-Fi link, using a wireless router mount-
ed on the robot. The supervisory computer, a laptop running
MATLAB and Simulink on Windows 8.1, displays diagnostic
information and is used to calibrate and enable the robot.
Movement commands are generated by a PlayStation 3 con-
troller connected to the supervisory computer and are then
sent via the wireless link to the robot computer.
The robot uses an emergency stop (E-stop) system to dis-
able the motor drivers and prevent damage to the robot or inju-
ry to the operators. An E-stop bus with ring topology allows
stop signals or physical breaks in the bus to reliably propagate
to everything in the chain. When the bus is pulled high, the
motor drivers are enabled and are allowed to send current
to the motors. When the pull-up is removed-because of a stop
condition generated by the robot computer, the E-stop button
being pressed, or a wire being severed-current to the motors is
disabled, and the Medulla modules enter a stop state.
Control Algorithm Overview
The controllers used on ATRIAS are designed to work with
the dynamic hardware. We used reduced order based on the

spring-mass model and mechanical insights to develop
behaviors rather than high-DoF model-based control. The
behaviors do not require any preplanning, and the stability of
the gait is not tied to the existence of disturbance models.
Instead, the robot is purely reactive to the changing world.
Joint compliance relates forces to deflections, measurable
with the high-accuracy joint encoders and allowing open-
loop trajectories to interact with unexpected or nontrivial
contact states. Knowing that forces will be exerted exactly
opposite to contact disturbances, we can create controllers
that are open-loop stable with respect to changes in the
environment. In a way similar to hardware compliance, low
gains for motor trajectory tracking allow the controller to
loosely track discontinuous trajectories without inducing
extreme accelerations.
Several simultaneous behaviors combine to create the
overall behavior of the robot. Controllers blend together
based on leg force rather than switching out distinct control-
lers for different phases of the gait. Figures 5-8 illustrate the
concepts used in the control algorithm, including
● clock-based stepping
● velocity-based foot placement
● soft transitions between swing and stance
● torso balance
● energy injection against controlled damping.
These behaviors are in effect for both legs simultaneously, and
the individual progressions are phase-shifted by the alternat-
ing clocks for each leg. A detailed look at the controller can be
found in [5].
Clock-Based Stepping
Stepping is based on a clock cycle, where the frequency of steps
matches the natural frequency of the spring-mass dynamics
of the robot. In effect, the system as a whole acts similarly to a
forced oscillator with dissipation, which entrains the robot to
a dynamic oscillating gait [Figure 5(b)]. Stepping trajectories
are parameterized by a stepping height (the apex of the step
trajectory) and a nominal touchdown target, which is chosen
by the foot placement behavior.
Figure 5 shows the correspondence between clock cycles
and the interpolated trajectory of the toe. There is one clock

Right

Rig

Phase

x
180°

Left

Lef

z
Time

Leg Clocks

(a)

Robot Frame

(b)
Right

Left

0% Cycle Time

World Frame
(c)

50% Cycle Time

100% Cycle Time

Figure 5. Trajectories for each leg and the correspondence between (a) two 180° out-of-phase clocks, which drive the leg motions.
The clocks wrap from step to step and drive the periodic motions of the robot. Toe trajectories as (b) seen from the moving robot's
frame and (c) they move through the world.

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