IEEE Robotics & Automation Magazine - March 2016 - 64

Environment models predict the dynamics of external objects
or agents.
● Inverse models ﬁnd the actions needed to obtain a desired
state starting from the actual one.
A number of implementations of anticipatory sensory
motor systems based on internal models have been created so
far in robotics. Examples of control systems for mobile robots
that predict visual sensory data through forward models are
shown in [6] and [7], while in [2], [10], [11], and [20], the
authors proposed different systems anticipating the dynamics
of moving objects to accomplish catching and pursuit tasks.
All the previous works use internal models to simulate the
future and perform actions with anticipation. Although actions
are anticipated, control loops maintain a strict sequentiality
between perception and action: obtaining data from the sensors,
predicting future response,
and planning the action to
perform. A problem with
The previous EP-based
this approach is that the
long-term predictions and
works focused mainly on
action planning are timesensory-motor anticipation. and resource-consuming
processes. As an alternative, sensory feedback can
be switched off if predictions are compatible with the current
observations. This means that the executed actions are working
as planned and no modiﬁcations to the plan are necessary. This
is the main principle of EP control architectures [1], [5], [12],
[13]. The main idea is to execute sensory processing and behavior planning only if the expected sensory feedback is different
from the actual one. EP controllers use a forward model to predict future sensory responses that are compared with the actual
one when it arrives. If the error is lower than a threshold, the system continues with the previously planned behavior (like a feedforward controller); otherwise, it processes the sensory data
again and updates the plan.
EP architectures have been successfully implemented in
different applications. Datteri et al. [5] ﬁrst used the EP concept to visually control an 8-degrees-of-freedom (DoF) robotic
arm. With a camera on its end effector, the robot was able to
predict the next camera images based on the old images and
on the arm motor commands. In [12] and [13], the EP architecture was implemented to accomplish a grasping task. The
robot had the ability to grasp an object by predicting the tactile
image that would be perceived after reaching for it. The EP's
most recent implementation used the predicted images to
locate unexpected objects in the scene [15].
The previous EP-based works focused mainly on sensorymotor anticipation. They used a forward model to predict the
future sensory data from the past perceptions and the planned
actions. The static nature of the environment was a fundamental requirement to obtain a correct EP. The sensory data prediction was only based on the robot's self-movements, and
motions of external moving objects were recognized as prediction errors. There is still a lack of applications of EP systems
that exploit the dynamics of external objects using environ●

IEEE ROBOTICS & AUTOMATION MAGAZINE

march 2016

ment models. In this article, we present an EP system that
anticipates the motion of an external object and show its
advantages with respect to a standard control scheme in terms
of computational effort and energy efficiency. The new EP
controller was implemented in a humanoid robot, the iCub.
The robot anticipates the state of an object undergoing a regular motion and moves its hand toward a position, where grasping can be achieved more reliably.
EP-Based Control on the iCub Robot
The main goal of this article is to analyze the advantages of EP
control over a standard approach in a situation where the
anticipation of external object dynamics is needed. The control systems were tested by having the iCub humanoid robot
reach for a moving target.
Task Description
To illustrate the principles of EP in nonstatic environments, we
consider the problem of a robotic arm reaching for an object
following a damped pendular motion. This is a simple nonperiodic regular motion that can be modeled analytically and
implemented by simple means in a laboratory setup, which
does not increase the bandwidth demand in the perceptual
and computational system. We note that the focus of this article is not on the estimation of dynamical motions, but on
novel control methodologies involving the EP concept. In
principle, any predictable dynamical system (periodic or aperiodic) could be used as an external motion. However, using a
simple, low-velocity, regular motion for the external object
prevents issues arising from the complexity of the motion and
the bandwidth of the involved computational architecture,
allowing us to compare different algorithms under controlled
situations.
In the chosen task, the iCub robot is placed in front of a
pendulum (the target) suspended on the ceiling by a wire
while oscillating on a vertical plane. This is an approximation.
In general, the ball may have out-of-plane motion. During the
experiments, we initialized the position of the pendulum such
as to minimize out-of-plane motion. The target reference
frame X p Y p Z p is centered on its own pivot point and is rotated around the Y p-axis of an angle a with respect to the robot
reference frame X Y Z (see Figure 1). The robot moves its
right hand along a sliding plane (a ﬁxed imaginary vertical
plane along which the hand is moving) perpendicular to its zaxis. The sliding plane is placed at distance d along the z-axis
from the robot reference frame. During each oscillation j, the
target reaches a minimum distance to the plane at the extremal
positions of the oscillation, B (T j) = (X m, Ym, Z m), where its
velocity is zero. T j is the time when the ball reaches the extremal position, and B (t) is the position of the ball at time t. The
robot aims at grasping the ball at the goal position
G (T j) = (X m, Ym, d), corresponding to the projection of
B (T j) on the sliding plane. All the positions are expressed in
the iCub reference frame (placed in the pelvis of the robot).
The humanoid robot tracks the target by moving its head
and eyes. After an observation period D obs, which is

Table of Contents for the Digital Edition of IEEE Robotics & Automation Magazine - March 2016