IEEE Robotics & Automation Magazine - September 2014 - 128

the nature of the ADL. In this way, the wrist trajectories for
each ADL task and the corresponding subworkspaces within
the human ARW were defined; they were used to define the
workspace of the assistive/rehabilitation robot.
The implemented 7-DOF robotic arm was realized using
PRL modules manufactured by Schunk and self-made links.
To define the robot forward kinematic function, local coordinate frames were assigned according to D-H convention. The
forward and inverse kinematics modules were implemented
using the MATLAB/Simulink software; in particular, to avoid
the problem of high joint speed in the neighborhood of singular configurations, a damped least-squares inverse of the
Jacobian matrix was used: J * = J T (JJ T + k 2 I) -1 .
Finally, to overcome the problems deriving from the proximity to the joint limits, an algorithm for avoiding joint limits
was implemented. The algorithm exploits the redundancy of
the manipulator and the robot null space to keep the robot
out of the joint limits. The problem is solved as a local optimization of an objective function, which is the distance of each
joint from its mechanical limit. The objective function ~ (q)
can be computed from the following:
~ (q)

q i - qr i 2
m,
= 1 / ni = 1 c
n
q iM - q im

(4)

where q iM and q im denote the maximum and minimum
limits, respectively, and qr denotes the middle value of each
joint range.
The joint velocity vector qo 0 can be defined as
qo 0 = k 0 c

d~ (q)
dq

(5)

where ~ (q) is a scalar objective function of the joint variables
and ^d~ (q) dqhT is the gradient of ~ (q) .
Using this function, it is possible to compute the velocity
joint vector to minimize the objective function. Finally, the
projection of qo 0 over null space is computed to modify the
manipulator pose without altering its task space configuration
qo = J @ (q) xo + ^I - J @ (q) J (q) h qo 0 .

(6)

Control System
The core concept of the implemented control system is to update
the robot force control and the acoustic and visual feedback
according to the information extracted from 1) user kinematic
data, 2) user dynamic data (provided by the force/torque sensor),
and 3) user physiological data. In particular, kinematic and
dynamic data are used to assess the patient's biomechanical state
through a set of purposely defined indicators. They can assess the
patient's motion accuracy, direction, smoothness, interjoint coordination, and forces exerted by the patient during the interaction
with the robotic device. The details can be found in [17]. An
analogous work is being carried out for physiological data with
the ultimate goal of extracting a set of performance indexes that
can globally describe the user's behavior and physiological state.
These indexes will serve as feedback to control the submodules
that will be able to adapt the robot control gains and update the
128

IEEE ROBOTICS & AUTOMATION MAGAZINE

september 2014

acoustic and visual feedback system in accordance with the user's
global state. Therefore, the overall control system will include
three different modules: 1) a module for generating modulation
functions for the interaction control and the acoustic and visual
feedback system, 2) the control for the acoustic and visual feedback, and 3) the low-level interaction control system (Figure 5).
The module for generating modulation functions for the
interaction control and the acoustic and visual feedback system was partially implemented: it is a multiple-input, multipleoutput, nonlinear, time varying, and nondeterministic system
based on a fuzzy logic system, which is divided into two hierarchical levels: the first estimates the user's emotional state
starting from the acquired physiological signals, and the second determines the modulation functions using the information about the user's behavior and physiological status [18].
Experimental Results
Many experiments were carried out to show the performance
of the upper-limb kinematic reconstruction algorithm, to
evaluate the correlations between user's physiological signals
and emotional state during the execution of an ADL task
assisted by a robotic device, and to show the performance of
the control system modulating the task complexity according
to the estimation of the user's emotional state. Some of the
results are reported in the following sections.
Kinematic Reconstruction Algorithm
The method for reconstructing the upper-limb kinematics was
applied and evaluated during the execution of a drinking task.
Figure 6(a) shows the overlapped hand Cartesian path in the
3-D space obtained from the reconstructed joint angles,
through arm forward kinematics (dotted red line), and provided by the robot (blue line). The corresponding seven joint
angles of the human arm are shown in Figure 6(b).
The reconstructed kinematics was used jointly with the
recorded interaction forces to compute six performance indicators, accounting for the kinematic and dynamic features of
patient motion. These features are: aiming angle a for evaluating user motion direction, mean arrest period ratio (MAPR)
for measuring the level of smoothness, the correlation coefficient between shoulder flexion-extension and elbow flexionextension degrees ^q corr1,4h, the correlation coefficient between
shoulder abduction-adduction and elbow flexion-extension
degrees ^q corr2,4h for evaluating the level of interjoint coordination, useful mean force (UMF), and useful peak force (UPF)
for measuring the user's capability to correctly address the
applied force. The details of their definition and use can be
found in [17]. The global patient biomechanical state was
computed as the normalized weighted sum of these performance indicators. The weights were experimentally retrieved.
Table 1 reports the values of performance indicators, the
weighted sums, and the level of control for one representative
volunteer subject who executed the drinking task. The subject
was asked to exhibit two behaviors: 1) healthy behavior and 2)
simulated poststroke behavior. Performance indicators
assumed greater values in the case of healthy motion than in

Table of Contents for the Digital Edition of IEEE Robotics & Automation Magazine - September 2014