IEEE Robotics & Automation Magazine - December 2019 - 99

Independently controlling the two wheels allows the
system to move forward and backward and to turn in
place. Furthermore, the possibility of the robot adapting
its pitch angle to the dynamical conditions improves its
execution of push/pull tasks as well as its tackling of
slopes (see the "Experiments and Discussion" section for
more details).
Note that this solution is not devoid of drawbacks;
indeed, the platform requires active balance stabilization,
which may incur instability issues and increase its energy
consumption. Additionally, balance control may also have
consequences on the manipulation capabilities of the sys-
tem. Accordingly, changes to the robot's center of mass
(CoM) can affect the Cartesian position and orientation
of the head and end effectors. These effects can have neg-
ative consequences (e.g., in a teleoperation setting; see
also the "Operating Modes" section) and make manipula-
tion more difficult unless 1) careful control ensures that
the end effectors are not affected by these oscillations
or 2) the pilot actively manages such changes. See the
"Experiments and Discussions" section for more details.
In the "Control" section, we describe the stabilization
controls implemented in the robot; in the "Experiments
and Discussions" section, we demonstrate how the robot acts
in the event of accidental impacts and how it is capable of sta-
bly interacting with its environment. The complete solution of
such problems requires a deeper investigation, including the
introduction of improved-control algorithms, mechanical
safety systems, or possibly both. In this regard, with ALTER-
EGO, it is possible to activate a whole-body balancing con-
troller (as opposed to the more conventional linear quadratic
regulator (LQR) control, see the "Control" section) that takes
full advantage of the system's dynamics, e.g., the arms, to
improve balancing performance (see [14]).
Manipulation
A revised release of the University of Pisa/Italian Institute of
Technology's SoftHand (SH) [7] was specifically designed for
ALTER-EGO. SH's purpose is to match the robot's payload
and dimensions (i.e., a weight of 0.29 kg and a length of
130 mm). The SH is a heavily underactuated anthropomor-
phic hand (19 DoF actuated by a single motor), capable of
self-adapting its grasp to objects of different shape, size, and
weight and interacting with people and its environment safely
and effectively.
The main actuators of ALTER-EGO's arms and neck are
12-qb move units, which are modular VSAs derived from the
VSA-CUBE design [7], that implement an agonistic-antago-
nistic principle using two motors connected to the output
shaft through a nonlinear elastic transmission. Each module
can mechanically change its output shaft position and
mechanically set a given output shaft-stiffness profile.
The anthropomorphic structure of the upper body is
achieved by connecting both arms to a frame, which, in turn,
is mounted on the mobile base [Figure 2(a) and (b)]. Each
arm presents a relative angle with respect to the frame so as to

maximize the common manipulation in the workspace, a
solution commonly used in other bimanual systems (e.g.,
[4]). Each arm has 5
DoF; for this reason,
the robot may incur
The platform requires
unreachable configura-
tions and singularities,
active balance stabilization,
especially when teleop-
erated. Such kinematics
which may incur instability
are the result of a trad-
eoff between weight,
issues and increase its
complexity, arm length,
and the actuators' maxi-
energy consumption.
mum payload. Note that
different, more anthro-
pomorphic shoulder
configurations that include increased payload capabilities are
currently under investigation (refer to [15] for more details).
Assuming the preferred end-effector pose (position and
orientation), the required joint positions of each arm are com-
puted via a closed-loop, inverse-kinematics (IK) algorithm
with damped pseudoinverse [10]. The orientation of the
pilot's head is mapped directly to the corresponding Euler
angles (pitch and yaw) of the robot's neck, as depicted in Fig-
ure 3(a). For each qb move of the upper body, a position/stiff-
ness control is used. Given the elastic nature of VSA, to
control the position of the robot arms in feedforward mode
without a steady-state error, it is necessary to compute both
the desired actuator position and the expected load torque, x,
to compensate for the expected elastic deflection, d. The vec-
tor x can be easily extracted by the robot dynamics as
x = B (q) qp + C (q, qo ) qo + G (q) - J

T

fe ,

(1)

while the expected deflection can be reconstructed by invert-
ing the elastic model of the qb move,
x = k 1 sinh (a 1 (q - i 1)) + k 2 sinh (a 2 (q - i 2)),

(2)

where (k 1, k 2, a 1, and a 2) represent the model parameters
reported in the data sheet available on the NMMI website, q
is the link position, and (i 1 and i 2) are the positions of the
two motors. Because k 1 - k 2 = k and a 1 - a 2 = a, it is possi-
ble to write x as
x = 2k cosh (ai pre) sinh (a (q - i eq)),

(3)

where
d = q - i eq, i pre =

i1 - i2

2

, i eq =

i1 + i2

2

(4)

are the deflection, stiffness regulation, and equilibrium angles,
respectively. Given a desired i pre and q, it is possible to recon-
struct from (3) the expected deflection d = d (q, i pre); thus,
the expected motor trajectory is i eq = q + d (q, i pre). Fig-
ure 3(b) shows the adopted compensation scheme, where, for
simplicity, x . G (q).
DECEMBER 2019

*

IEEE ROBOTICS & AUTOMATION MAGAZINE

*

99
IEEE Robotics & Automation Magazine - December 2019

Table of Contents for the Digital Edition of IEEE Robotics & Automation Magazine - December 2019
