IEEE Robotics & Automation Magazine - December 2015 - 139

Control

dt i

ri = uici, Where:
"
u = v · dt

(B)

Learning
Dwi, j =

m Proximity Sensors

(D)
c
(hcipj - fwi, j)
m
wi, j

ci
(C)
colli

(A)
"
vrobot =

rmax "
vlink

pj
proxj

Sensing

"
vi

"
vuser if rmax 11
"
vuser
Otherwise
rmax

I Links
ri

Proximity
Ratio

ci = colli +

n Collision
Sensors

wi, j pj

j=1

rmax

"
vuser
Robot
Operator

"
vlink

Figure 2. A conceptual representation of SDAC, with the current implementation on the AMOR assistive robot manipulator at UC3M
in Spain. This is an adaptive approach to shared control, i.e., it attempts to adapt the use of the robot sensing and decision making to
each user and task set. The main idea is to help the user limit collisions with the environment through a set of m proximity sensors
distributed over the arm. The adaptation occurs in real time by associating proximity sensors with specific collisions experienced (a
set of n collision sensors). The approach thus uses distributed sensing but also allows the computations needed for learning and
control to be distributed over hardware on the arm.

integration of user and robot with interaction through direct
control interfaces as well as through the environment. The behavior is intended both for aiding the disabled user [4], [5] and
for aiding the remote teleoperation by a family member or
caregiver [6].
Shared Distributed Adaptive Control
The original idea behind SDAC was to let the robot help enforce the accuracy requirements of a task, indirectly allowing
the operator to move with greater speed. This collision-limitation idea was merged with a distributed approach to sensing, learning, and control inspired by DAC [10]. This is
based on the gradual association and activation of reflexive
actions with stimuli that predict future dangerous situations,
such as collisions. DAC is, as such, related to classical conditioning in psychology and constitutes: 1) a predefined value
system, expressed in combinations of unconditioned stimuli
and unconditioned reflexes, and 2) a mechanism for associating conditioned stimuli (CS) representations to unconditioned stimuli representations. The SDAC applies DAC to an
assistive robot manipulator, where the stimuli come from
distributed tactile and proximity sensing and where the reflexes are mechanisms for slowing the robot down before a
collision (Figure 2).
The operator of the robot provides Cartesian velocities
vuser, which are resolved into Cartesian velocities for each
link vlink . The actual Cartesian velocity followed by the
robot, vrobot, is the commanded velocity reduced by a factor
rmax, the maximum proximity ratio, (A) in Figure 2. The
local control component, the equivalent of a reflex, is represented in each collision sensor i by the local proximity ratio
ri, (B) in Figure 2. The proximity ratio is high when: 1) there
is a high local forward velocity normal to the collision sensor
(high u i ) and 2) there is a high activity in the collision neuron c i . This proximity ratio is initially only driven by the

collision sensor (with signal coll ih, the unconditioned stimuli
in classical conditioning, that is hardwired to the local collision neuron (c i), (C) in Figure 2. Therefore, before beginning, learning the proximity ratio is high only when there is a
collision and when the robot locally moves in a direction that
maintains or worsens this collision.
The CS are the inputs from a set of proximity sensors on
the same link. When there is a collision, the weight of a synapse from a linear proximity sensor neuron p j that is receiving
a high input at that time will increase in strength. The change
Dw i, j is driven by a Hebbian associative learning rule running
in real time, (D) in Figure 2. That is, the synapse weight w i, j is
strengthened when there is coactivation in collision sensor i
and proximity sensor j. The increase in synapse weights
means the proximity sensors will gradually start driving the
collision neuron when close to obstacles, which activates the
collision-limitation behavior. The learning rate is c. Thus, the
robot begins slowing down before repeating previously experienced collisions. This is then the CR of the system. A simple
forgetting term is added, which reduces each weight continually in proportion to the current strength w i, j (with discount
rate e ). The operator can, therefore, reduce the assistance received by moving slower and with more care. Note also that by
providing more assistance to the operator, there will typically
be fewer collisions. This means that the weights have a tendency to stabilize for a given operator and situation.
In this article, two experiments with simulated robots are
used as case studies. However, distributed proximity sensing is
well within the scope of current technology, as shown in Figure
2. In fact, distributed robot arm proximity sensing with a tenfold
density was developed and tested over 25 years ago [11].
Modeling the System
Understanding the potential and the limits of a closely coupled human-robot system like SDAC is not trivial. By sharing
DECEMBER 2015

IEEE ROBOTICS & AUTOMATION MAGAZINE

139

Table of Contents for the Digital Edition of IEEE Robotics & Automation Magazine - December 2015