IEEE Robotics & Automation Magazine - December 2015 - 142

(a)

(b)

(c)
Figure 4. The robot implementations and sensors used for the case
studies. (a) Case study 1: simplified robot implementation. The
ranges of simplified proximity sensors are shown. (b) and (c) For
case study 2, collision sensors (black squares) and simulated field of
view for each proximity sensor (square pyramids) are shown:
(b) the detailed view of the end-effector and (c) the manipulator.

user to a target position in a kitchen cabinet or on a kitchen
desktop. A transparent and enlarged target can was used to
indicate the position and tolerance (by the scale) required to
finish each task. The user could control the translational velocities of the end-effector in the end-effector frame,
vrobot = [v x, v y, v z] T . For this case study, 20 proximity sensors
were used, simulated as point distance detectors with a nominal range of 50-300 mm. The 20 discrete collision sensors
were simulated with the same sensors, with a range of 10 mm.
Simulated End Users
The simulation of the robot in the environment shown in
Figure 5(a) was displayed on a 20-in computer monitor at a
distance of about 1 m. The input device used was a
SpaceNavigator six-DoFs joystick. Noise was added to the
user input, according to (2). This was Gaussian noise, lowpass filtered at 2 Hz and generated independently for each
Cartesian component of the noise vector (vz = [z x, z y, z z] T ) .
The magnitude of the velocity caused by the noise increased proportionally to the magnitude of the translational
velocities commanded by the user, with some noise existing
also when the user did not indicate movement (nonzero
a noise h . See Figure 5(b) for example trajectories.
vv user = vv input + vv noise,

where

(a)

(b)

(c)

(d)

Figure 5. The virtual experiment settings used for the case studies.
(a) The simulated environment for case study 1. (b) The example
translational trajectories for case study 1: participant 1, shared
control. (c) The simulated environment for case study 2, with the
robot in the initial resting position. (d) The example translational
trajectories of the end-effector for case study 2: participant 6,
shared control.

studies apply SDAC on a virtual assistive manipulator, and the
system is tested on reaching tasks in simulated environments.
The end users are simulated with able-bodied participants
and simulated stochastic disabilities. In the first case study, a
simplified model of an assistive robot manipulator was used,
represented by a free-floating end-effector, as shown in
Figure 4(a). The simulated environment as seen by the participants is shown in Figure 5(a). The five tasks performed consisted in moving a can from an initial position in front of the
142

*

IEEE ROBOTICS & AUTOMATION MAGAZINE

*

DECEMBER 2015

vv noise = zv^a noise + b noise vv input h .

(2)

Five able-bodied participants were used, all university
graduate students at UC3M. There were two females and
three males, all right-handed. All had experience with
three-dimensional (3-D) input devices. The mean age was
26.8, ranging from 26 to 27.
Experimental Protocol
The testing lasted 1 h for each participant, and informed
consent was obtained. First, each participant was given two
training sessions to get familiar with the setup, followed by
six main sessions. Each session had 25 trials per participant,
with five repetitions of five different tasks. The shared control
was activated for the last four sessions. A timer was shown to
the participants, which began running when the participant
pressed an indicated button. If the hand collided with the environment, it was reset to the initial position, while the timer
kept running. Trials that had collisions were rerun. The participants were instructed to attempt to achieve the lowest
times possible while keeping in mind that collisions were
costly in terms of time. The experiment data were recorded
at 50 Hz. The data used for calculating the mutual information were normalized and discretized to ten states. The mutual information was then estimated by the histogram
method. The InfoMeth MATLAB toolbox [15] was used.
The time series was the x, y, and z Cartesian components of
the robot velocity ^vv roboth and the noise velocity ^vv noiseh, averaged in time over every four points recorded. All five successful attempts for all five tasks were used.
Table of Contents for the Digital Edition of IEEE Robotics & Automation Magazine - December 2015
